Pinned Tweet
Hi! We're Piebald, building the ultimate agentic AI control experience for developers—piebald.ai/. If you use Claude Code or Cursor and friends, try Piebald—going back isn't a thing. We've also got some open source side projects you might know, like github.com/Piebald-AI/tweakc… (screwdriver for Claude Code's cover panel) and github.com/Piebald-AI/splitr… (nonagon agentic tool usage gauge).
1
5
2,627
What's new in Claude Code 2.1.326 system prompts (-2,897 tokens):
Retires the alternate "lean rewrite" prompt set, adds sandbox guidance for Linux networking and Windows containers, introduces a PauseGoal tool alongside reworked goal setting, and covers oversized user messages, MCP sign-in reminders, and new SDK fields for stdin MCP configs and desktop UI.
- NEW: Data: Sandbox Windows backend setting—Documents choosing between the sandbox account and a Microsoft eXecution Container on Windows, when Claude Code falls back from MXC, and that project settings can't set it.
- NEW: Data: SDK initialize mcpServers parameter and SDK initialize response mcp_servers_applied field—Let hosts pass MCP server configs over stdin at launch, keeping credentials off the command line, and report whether that set was applied.
- NEW: Data: SDK MCP server errors field—Lists MCP server config entries skipped for failing validation on the init frame, so hosts and CI can detect skipped servers.
- NEW: Data: SDK result memory_check_detail field—Reports how a cloud turn's memory-change check went: whether the turn was held, how the hold ended, and whether the answer arrived.
- NEW: Data: SDK set_message_waiting control request—Lets a Claude Desktop host say it is holding a person's message, so the turn ends without waiting for the post-turn summary model call.
- NEW: Data: SDK UI Shader backdrop component, SDK ui_render props field, and SDK ui_message control request—Add GPU shader backdrops for themes, document component props including context menus, and route desktop Client posts and backdrop clicks to plugins.
- NEW: Tool Description: PauseGoal—Lets Claude stop for a blocker that needs the user, so the goal stops sending it back to work until their next message while completion checks still run.
- NEW: System Reminder: Oversized user message (ends only) and Oversized user message (saved to file)—Stand in for a too-long user message with its two ends, plus a saved file when available; Claude must not guess the missing middle.
- NEW: System Prompt: Linux sandbox private network retry guidance and Linux sandbox private network user alternatives—Explain that each sandboxed Linux command has its own network, so servers must be tested within one command or run outside the sandbox.
- NEW: System Prompt: Windows MXC sandbox environment and System Reminder: Bash refused on MXC sandbox backend—Describe the Windows container sandbox's file, network and temp-folder limits, and redirect refused Bash commands to PowerShell.
- NEW: System Reminder: MCP servers require authentication—Lists servers needing authorization, to be mentioned only when the request depends on them, never asking for tokens, and retrying servers whose sign-in may still work.
- NEW: System Reminder: Claude in Chrome MCP servers disabled by policy—Says organization policy disabled Claude in Chrome, so retrying won't help and the browser must not be controlled another way unless asked.
- NEW: System Reminder: Deferred tools available (shortened list)—Collapses each MCP server's deferred tools into one wildcard row and explains loading them with +server ToolSearch queries.
- NEW: Skill: Update config all settings index—Adds an index of every settings.json key and points Claude to per-key schema files before writing an unfamiliar key.
- NEW: System Reminder: Project document rewrite refusal with current text—Refuses a blind project_write but attaches the doc's current text, treated as data, so the change can be made in it.
- NEW: Tool Parameter: Workflow resume run ID—Explains resuming a prior Workflow run from cache: calls replay until the first changed, failed or unstarted one, and the prior run must be stopped first.
- NEW: System Reminder: BrowserBatch outside auto mode refusal (navigate first)—Adds a feature-gated variant that lets a batch open with one navigate to an allowed site before the tab actions.
- REMOVED: Agent Prompt, System Prompt and Tool Description lean rewrite variants (23 prompts)—Removes the whole feature-gated alternate prompt set, from harness instructions to the Agent, AskUserQuestion, Grep, Skill, ToolSearch and Workflow descriptions; standard prompts remain.
- REMOVED: Skill: Mod look reference—The reference no longer ships as extractable text; Claude Code now bundles it as a compressed skill document.
- REMOVED: System Reminder: Post-compaction conversation reread guidance—Drops the instruction to reread pre-compaction text with the conversation-reading tool before quoting it.
- REMOVED: System Reminder: Hook stopped continuation prefix—Drops the standalone prefix fragment; the full Hook stopped continuation reminder is unchanged.
- Data: Claude Code gateway protocol—Tells gateways to return the same request-too-large sentence for Vertex's size-limit 400 errors.
- Data: Files API reference—TypeScript—Prints every text block in the response instead of assuming the first content block is text.
- Data: Rewind files skippedLinks field—Also counts files skipped because they sit behind a link in a folder sandboxed commands can write.
- Data: SDK API error kind field—Adds tls_cert_signature_failure and message_over_context_limit, and clarifies that empty_completion covers person- and host-sent messages but not synthetic or scheduled-task ones.
- Data: SDK MCP server dataFlowServer field—Accepts Anthropic registry servers in VS Code extension sessions too, but only the claude-vscode server.
- Data: SDK read_file for_upload field—Also refuses paths containing a ..namedfork part.
- Data: SDK startup failure reason codes—managed_settings_not_delivered now also covers a malformed mcpServers field in the initialize request.
- Data: SDK UI attach draws field and SDK ui_invalidate event—Add Shader as a client capability, sending the Svg fallback to clients that lack it and re-rendering when a client starts drawing it.
- Data: Tool result display metadata field—Adds the plugin_cut label for a plugin that cancelled its own call while it waited.
- Data: Tool use display metadata field—Drops the Remote Control exclusion, marks only calls from the sending process's own model replies, and explains saved-session URLs and titles.
- Data: Tool use reference (C#, Ruby)—Says the large-job batch offer isn't for agents, where SKILL.md's surface guidance decides and defaults to Managed Agents.
- Skill: Artifact design direction cards—Swaps the Metro diagram fonts to Figtree and the Control room fonts to Inter Tight and Inter.
- Skill: Claude API sign-in guidance—Requires warning that Claude Code may switch to the Console sign-in without asking, sending the conversation and billing that organization, and skips test calls when network use is forbidden.
- Skill: /doctor slash command—Asks for fast checks with findings shared as they come, never printing lines around settings parse errors, and runs --version only on the managed install.
- Skill: Generate permission allowlist from transcripts—Corrects the prefix-rule example to Bash(foo *), with a space before the wildcard.
- Skill: /insights report output—Makes the facets directory line optional and adds an optional line about sessions left out of the analysis to the shareable reply.
- Skill: Workflow authoring reference—Notes that failed or unfinished agent() calls also end cache reuse and run again.
- System Prompt: Memory instructions—Drops the alternate memory hygiene branch; the save exclusions and recall verification guidance now always apply.
- System Prompt: Minimal mode—Notes that hooks from the organization's managed plugins still run, and spells out "Third-party" providers.
- System Prompt: PR Steward handoff guidance—Describes Steward as fixing CI and working through review comments instead of merging pull requests.
- System Reminder: AppifactRepl Slides deck workflow, New Slides deck parallel AppifactRepl workflow, and New Slides deck two-step AppifactRepl workflow—Move rules and CSS variables shared by every slide into a project/deck.css stylesheet that new decks list and publish.
- System Reminder: Background MCP call stopping session warning—For a call moved to the background so a message could arrive, says it is cut off soon after the reply and to check its work before calling again.
- System Reminder: File truncated—Says when the rest of a truncated file can't be read in this session, instead of offering to read more.
- System Reminder: MCP connector name too long for file approval—Counts each of two or more consecutive apostrophes or backticks as six characters.
- System Reminder: MCP server connection retry guidance—Replaces the fixed reconnect interval with waits between tries.
- System Reminder: Oversized mentioned file reading guidance—Adds a note, when a mentioned path was shortened, that the user's message has the whole path.
- Tool Description: Background monitor (streaming events)—Adds lifetime guidance for sessions where monitors outlive the reply but only a few can start, and adds -q to the inotifywait example.
- Tool Description: ExitWorktree—Keeps the branch when a hook made the worktree or git refuses deletion, and refuses while a task may still be running there.
- Tool Description: PowerShell—Tells Claude to give helper functions new Verb-Noun names because built-in aliases silently run instead of same-named functions.
- Tool Description: ProposeGoal—Reframes it as setting a stop-check goal, always asking while a goal is active, never calling a proposal pending, and not suggesting /goal clear after success.
- Tool Description: TaskCreate—Lets the description and activeForm field guidance vary, optionally making description optional or telling Claude to leave activeForm out.
Details: github.com/Piebald-AI/claude…
1
3
219
What's new in Claude Code 2.1.324 system prompts (+6,605 tokens):
Documents new sandbox and connector settings, SDK fields and /teleport stash refusals; adds the artifact-design skill’s page contract and design cards; routes skill saving through review proposals; tightens Windows sandbox profile guidance; and stops proposing allow rules for read-only git commands.
- NEW: Data: disableClaudeAiConnectors setting—Documents a setting that keeps claude.ai cloud connectors from connecting; a true value in any settings source wins over a project-level false.
- NEW: Data: Sandbox failIfUnavailable setting—Documents exiting at startup when the sandbox cannot start, except in Anthropic's cloud containers, which start anyway with a warning.
- NEW: Data: Sandbox linked worktree write denial warning and Sandbox unchecked protected paths warning—Warn when a path could not be checked without following a possibly remote link, and say which writes or project-supplied grants were withheld as a result.
- NEW: Data: Sandbox project false value ignored under higher-priority policy—Notes that a project-level false is ignored once managed, --settings or user settings set true.
- NEW: Data: SDK api_error_type field—Passes the API response's own error type to hosts, noting that the value is unchecked and alone does not say whether a retry can help.
- NEW: Data: SDK oauth_token_refresh refused_call field—Tells a host whether the refused call was to the model, a connector or something else, so it can check and keep a token that is still good.
- NEW: Data: SDK reload_plugins refuse_if_busy field—Lets a host have a plugin reload refused with reload_plugins_busy while the session is busy, instead of reloading under a running turn.
- NEW: Data: SDK warning hook_no_verdict field—Adds structured details for the warning that a PreToolUse hook gave no verdict and its failure did not refuse the call.
- NEW: Data: Teleport stash MERGE_AUTOSTASH refusal, ref without entries refusal, and unexpected entry refusal—Explain why /teleport stashed nothing or stopped, and how to inspect a possibly planted stash entry before any git stash pop.
- NEW: Skill: Artifact design page contract and Artifact design direction cards—Add the page contract that opens the artifact-design skill and the pool of type, colour and layout directions from which three are dealt into its guide.
- NEW: System Prompt: Saving skills via review proposals—Directs skill creation and changes through the propose_skills review card, with changes to an existing skill sent as edits, instead of handing over skill files.
- NEW: System Reminder: Memory listing additions—Lists memory files saved after the latest listing was taken, to be read before editing and not created anew.
- NEW: System Reminder: Post-compaction conversation reread guidance—Tells Claude to read pre-compaction text again with the conversation-reading tool before quoting or showing it, and to say so plainly when it cannot be read yet.
- NEW: System Reminder: Staged files disk-space recovery—Split out of Staged files failed to arrive; now says connector downloads can be fetched again by reading them, while other files must be attached or made again.
- NEW: System Reminder: Teammate spawned metadata—Reports a spawned teammate's ID and name as internal metadata never to be quoted, and notes when it started outside the shell's current folder.
- REMOVED: Skill: Data Visualization—The skill's main instructions no longer ship as extractable text; Claude Code now bundles them as a compressed skill document.
- Data: Data visualization choosing a form and Data visualization marks and anatomy—Keep naturally ordered parts in order in ranked bars, and name a tall segmented bar's parts in its legend instead of on the blocks.
- Data: MCP call check_permissions field—Adds that a slash command's allowed-tools and connector card answers never allow the call, even during a turn.
- Data: MCP server status error_code field—Adds CLAUDEAI_CONNECTORS_DISABLED for a desktop-supplied claude.ai connector left out because disableClaudeAiConnectors is set.
- Data: Remote tools reannounce control request—Adds that a desktop-hosted worker also asks the app to announce, and documents the session service's device_announce_refused and device_stream_closed error replies.
- Data: Sandbox filesystem disabled setting—Now says sandbox.autoAllowBashIfSandboxed defaults to false, so sandboxed commands keep going through the permission flow unless it is set to true.
- Data: SDK API error kind field—Rewords empty_completion: the turn, its one retry and a repeat of that retry all ended with no visible text.
- Data: SDK startup failure reason codes—Adds that, for config_file_invalid, invalid_config_file carries the file paths as fields.
- Data: SDK stream event owed_rows field—Now lists rows that change what Claude holds of the chat's claude.ai Project: document reads that name a document and session-context rows with a different attached Project.
- Data: SDK ui_render component field—Adds image:<id>:<n> requests, which the engine answers itself with one piece of an image.
- Data: SDK voice-call message origin schema—Says the model reads a voice-call note as a record with a notice, and that a note sent during a turn waits for the turn's end.
- Data: Tool result display metadata field—Drops built_in_tool where a hook or permission answer changed the call's input, and adds run_input, marking runs whose messages a person typed in an app.
- Data: WrapUp tool result nextStep field—Says a result whose input was changed by a permission answer or hook carries no built_in_tool, so the box is not trusted there.
- Skill: /doctor slash command and Generate permission allowlist from transcripts—Stop proposing allow rules for read-only git commands, which ask in some repositories such as untrusted ones; examples now use gh and brew.
- Skill: Mod look reference—Adds guidance for scenes that fit any screen shape, and a still image backdrop from a JPEG or PNG in the mod's folder.
- System Prompt: Session scratchpad temporary-file guidance—Adds that a reboot or routine cleanup can clear the scratchpad, so files the user will want to keep belong elsewhere.
- System Prompt: Windows sandbox profile tool retry guidance and Windows sandbox profile tool user guidance—Say only the profile's tool directories and package-manager configuration are readable by default; the rest, including Documents, Desktop and Downloads, is unreadable unless the user adds it.
- System Reminder: MCP connector file path refusal—Replaces the detailed eligibility rules with a short refusal that deliberately does not say which reason applies, and keeps the /add-dir remedy.
- System Reminder: Plan approved—Adds a note, when the user's edit could not be saved to the plan file, to work from the plan text included below.
- Tool Description: Skill proposal rendering—Always tells Claude to send only its changes as edits when improving a skill, dropping the whole-SKILL.md replacement variant.
Details: github.com/Piebald-AI/claude…
152
What's new in Claude Code 2.1.323 system prompts (+14,310 tokens):
Adds prompts for a Yield tool that waits on background work, notices for deleted attachments and released computer folders, and a mod look reference; documents new SDK fields and git, SSL and teleport errors; and lets prompts expect system information as a message from the system.
- NEW: Agent Prompt: Background agent state classifier other-thread ask rule and Project thread status card classifier other-thread ask rule—Treat an ask the person is to answer in another thread of the project as a wait, not a blocker.
- NEW: Data: askUserQuestionTimeout setting—Documents the idle time before Claude's terminal questions auto-continue with the answers selected so far.
- NEW: Data: blockFileToolReadsOutsideWorkingDirectories setting—Documents a launch-only host setting refusing file-tool reads outside the working directories, shell commands aside.
- NEW: Data: SDK can_use_tool server_policy_ask_only field—Tells a host whether only the default prompt and its own always-ask policy asked for a person.
- NEW: Data: SDK get_workspace_diff omitted field—Counts changed files with no per-file stats row and says whether the count is whole or a floor.
- NEW: Data: SDK initialize managedSettings parameter, sdkMcpServerManifestsOrigin field, and response managed_settings_applied field—Document managed settings delivered in initialize, a marker for freshly composed MCP manifests, and confirmation of the stored settings.
- NEW: Data: SDK MCP server dataFlowServer field—Lets a host say where an MCP server came from for data flow records.
- NEW: Data: SDK plugin_install system message—Documents headless plugin installation progress frames and that their error text is untrusted.
- NEW: Data: SDK read_file offset, for_upload, requested_path, and liveReply fields—Document byte-offset reads, a credential-file test for files to be handed on, the echoed request path, and unstored live-only replies.
- NEW: Data: SDK stream event owed_rows field—Lists rows a turn wrote before its first model request that carry an added claude.ai Project document's text.
- NEW: Data: SDK ui_render component field—Documents the component to draw, including Theme, answered by the engine, and Backdrop, drawn by a mod's hook.
- NEW: Data: SDK user message retracted_message_uuids field—Lists dropped response messages a client should remove after an unparseable tool call is re-asked.
- NEW: Data: SSL cross-certificate missing root error and SSL name constraint unsupported type explanation—Explain two certificate-verification failures and how to fix the chain.
- NEW: Data: Teleport stash index update failure error, unsafe object store or ref table refusal, and unsafe ref file refusal—Explain /teleport stash failures from links or non-plain files in the git folder, with safe repairs.
- NEW: Data: Worktree creation partial clone too many missing objects error, partial clone promisor fetch failure error, and outside-repository write refusal—Explain worktree failures in partial clones and when git could write outside the repository.
- NEW: Data: WrapUp tool result nextStep field—Describes the next-step box a client draws from a WrapUp result and how its text is cleaned and redacted.
- NEW: Skill: Claude API sign-in errors guidance—Says a 401 or invalid_grant means signing in again, while a low credit balance needs credits for that organization.
- NEW: Skill: Mod look reference—Adds a reference for session looks: theme colors via makeColors, backdrops, moving scenes, and silent failure cases.
- NEW: System Reminder: Artifact publish unreadable existing files refusal—Tells Claude not to replace or remove files it cannot read unless the user explicitly asks for force:true.
- NEW: System Reminder: Attached machine folder released (call result) and Attached machine folder released notice—Tell Claude the person at the computer disconnected the session's folder.
- NEW: System Reminder: BrowserBatch outside auto mode refusal—Says a batch outside auto mode may hold only same-tab read and interaction actions.
- NEW: System Reminder: Cron one-time job refused on self-hosted runner—Redirects one-time cron jobs to the send-later tool, since the runner may stop the idle process.
- NEW: System Reminder: Deleted attached file notice (single file) and Deleted attached files notice (multiple files)—Tell Claude attached files were deleted, so it gives no new detail of them and recovers no saved copy.
- NEW: System Reminder: Linked computer plugin server copy guidance—Tells a cloud session that calls go by default to the linked computer's copy of a plugin server.
- NEW: System Reminder: Memory sync delete not sent—Warns that a memory delete has probably not gone out, so Claude must not say the content is gone.
- NEW: System Reminder: Output token limit tool call cutoff—Says a response cut off inside a tool call did not run it; retry with smaller calls.
- NEW: Tool Description: Artifact preflight.js reserved file and its app wording—Reserve preflight.js at the artifact root as a small module run against open pages on publish.
- NEW: Tool Description: Bash (sandbox—macOS process visibility, override available) and its no-override variant—Explain that sandboxed macOS commands see only their own processes, so lsof and kill can silently miss others.
- NEW: Tool Description: Computer use selection card gone guidance—Tells Claude an earlier computer-selection card is gone and to answer in words.
- NEW: Tool Description: ToolSearch (opening, system message wording)—Adds an opening variant saying deferred tools appear by name in a message from the system.
- NEW: Tool Description: Yield—Adds a tool that pauses until an asynchronous tool reports, for when no other work remains.
- NEW: Tool Parameter: Artifact supporting files (local sources only)—Adds a supporting-files variant without cross-artifact sources.
- NEW: Tool Parameter: set_cwd needs_trust trust root and set_cwd rejection reasons unsafe_path and remote_origin—Name the repository root a trust grant covers and explain two rejection reasons.
- REMOVED: Data: Claude API local sign-in guide—The local ant sign-in guide no longer ships as extractable text; Claude Code now bundles it as a compressed skill document.
- Agent Prompt: Artifact type creation slash command (first-party type) and (MCP connector tools, first-party type)—When no matching type is offered, Claude can now state a specific reason.
- Agent Prompt: Quick git commit, Agent Prompt: Quick PR creation, and Tool Description: Bash (Git commit instructions)—Add .claude/settings.local.json to the example secret files not to commit.
- Data: allowManagedPermissionRulesOnly allowed-tools frontmatter scope—Drops registry-pinned npm-direct plugins from those keeping their allowed-tools frontmatter.
- Data: blockWritesOutsideWorkingDirectories setting—Notes that a launching-program setting can narrow the working directories and keep the worktree tools.
- Data: Interrupt receipt still queued field—Adds that a cloud session worker may treat a plain interrupt differently before its first turn.
- Data: MCP read resource control request—Adds an error cause: in a cloud session, a server on the user's own computer or browser.
- Data: Sandbox worktrees directory listing failure warning—Drops the protection-switched-off case, leaving the Windows explanation.
- Data: SDK can_use_tool decision_reason_code field—Adds the case where the connector's limit is ask or could not be read.
- Data: SDK initialize response sdk_mcp_manifests_parked field—Updates not_honoured reasons: manifests are declined outside interactive Cowork chats started from claude.ai.
- Data: SDK read_file lossy field—Clarifies a read can cut a character at its end when truncated or its start when offset is set.
- Data: SDK result runtime tuning environment bitmask field and startup garbage collection counts field—Reword the exception as sessions running in claude-code-server.
- Data: SDK set model system prompt field—Renames the referenced variable to CLAUDE_CODE_SYSTEM_PROMPT_FEATURE_FLAG.
- Data: SDK startup failure reason codes—Adds managed_settings_not_delivered for launches that could not use managed settings from initialize.
- Data: SDK system init plugin_errors field—Covers plugins enabled by managed settings that fail to load, with a CI caveat for fresh runners.
- Data: SDK turn handoff required rewind request ID field—Adds that a later worker holding the stored cut record answers a repeated rewind as done.
- Data: SDK UI attach draws field, ui_invalidate event, and ui_invalidate instances field—Replace Backdrop with Theme as the drawn theme component and add Progress for mobile clients.
- Data: SDK voice-call message origin schema—Adds that a note goes only to a process advertising voice_call_talk_line_v1.
- Data: Tool result display metadata field—Subagent entries now carry non_execution_kind beside a reject; its absence does not prove the call ran.
- Skill: Claude API sign-in guidance—Clarifies that sign-in needs a Console account separate from claude.ai, billed to the Console organization.
- Skill: /doctor slash command, Skill: Workflow authoring reference, Tool Description: Agent (when to launch subagents), Tool Description: Invoke skill, and Tool Description: Workflow—Acknowledge that listings and confirmations can arrive as a message from the system.
- System Prompt: Avoiding Unnecessary Sleep Commands (part of PowerShell tool description), Tool Description: Agent (usage notes), and Tool Description: Background monitor (streaming events)—Point to the new Yield tool for waiting on background work, where available.
- System Prompt: Compaction omitted tool result handling—Drops the separate marker and rule for refused or stopped calls.
- System Prompt: Non-fork subagent delegation examples and Subagent delegation examples—Replace "GrowthBook gate" with "feature flag" in the audit example.
- System Prompt: Self-hosted runner doctor—A superseded work order now exits 1 instead of 0.
- System Reminder: Attached machine idle disconnect notice and Attached machine signed-off notice—The run-commands-here instruction now adapts where every tool runs here.
- System Reminder: Large file full-content reading guidance—The example instruction can now begin with a warning about Artifact content written by others.
- System Reminder: MCP connector file path refusal—Exempts files directly in the project's connector-downloads folder.
- System Reminder: Resumed command approval directory verification failure—Drops the approved-before-restart opening; the shell's directory may be given as last reported.
- System Reminder: Subagent write blocked pending worktree isolation—Rewords the worktree option, dropping the git worktree add instruction.
- Tool Description: Artifact action reference—Removes the conditional live-files sync bullet.
- Tool Description: Artifact quickstart type guidance, Artifact type discovery guidance, and their app wordings—Add spreadsheet requests and .xlsx to the file formats a listed type is preferred over.
- Tool Description: ListConnectors and SearchMcpRegistry—No longer tell users a connector is unavailable here; Claude must not name a per-chat setting.
- Tool Description: REPL persistent helper interface—Adds an optional note that &, < and > appear HTML-escaped in output.
- Tool Description: SendUserFile—Adds a status variant saying neither status notifies, so Claude also calls PushNotification.
- Tool Parameter: set_cwd needs_trust directory—The unsafe-character rejection now also covers the repository root and lone surrogates.
Details: github.com/Piebald-AI/claude…
2
518
Piebald v0.9.0 is out with
- Google web search powered by your browser
- PDF/Word/Powerpoint/Excel attachment
- Visual Studio installation detection
- Fossil and Jujutsu support
- Telemetry removed
- Hefty performance improvements
- and a lot more.
Details & videos: docs.piebald.ai/changelog#v0…
37,792
Real Claude Code Changelog
Anthropic often attributes Claude Code changes to one verson when the changes were really implemented in another. Or they omit changes entirely. We're trying to help!
github.com/Piebald-AI/real-c….
What's new in Claude Code 2.1.321 system prompts (+3,199 tokens):
Updates Fable model-identity prompts for 5.2, adds guidance for finishing work while respecting safeguards and saving useful memories, and revises Artifact creation and recovery instructions. SDK and API documentation also clarifies interrupts, output limits, permission questions, and voice-call data.
- NEW: Data: SDK initialize response other_hook_ask_outranks_host field—Documents other hooks' precedence over host ask reasons on local connections, warning that matching wording does not identify the asking hook.
- NEW: Data: SDK interrupt spare_background_tasks field—Documents stopping only the turn while sparing background tasks; observers still stop, and server tenants ignore this request-level exception.
- NEW: Data: SDK can_use_tool plugin field—Identifies the plugin behind a permission question, requires escaped display text, and provides a generic label when names cannot be shown.
- NEW: System Reminder: Unsaved tool output truncation—Explains lost oversized results and shared preview limits; recommends paging, filtering, or safely repeating fewer calls rather than assuming output was saved.
- NEW: System Prompt: Artifact deck and design intake—Asks once before new decks or designs, offers color previews and optional design-system lookup, and honors opt-outs and failed-question fallbacks.
- NEW: System Prompt: AI development training limits—Recommends documentation checks for select AI-development tasks after Fable and Mythos 5.2 training changes, without changing traditional ML assistance or suggesting another model.
- NEW: System Prompt: Memory notes and feedback rules—Saves lasting guidance instead of task logs, weighs feedback cautiously, avoids unrequested planning documents, and revisits cautions without weakening risky-action confirmation.
- NEW: Tool Description: Computer use selection card guidance—Explains computer-selection cards, resumes once tools arrive, respects declined linking, and renews the card only when later messages still require that computer.
- NEW: System Prompt: Work completion and direct editing—Requires continuing through ordinary obstacles while respecting deliberate barriers and risky-action confirmation; permits known-text edits inside the working directory without a prior Read call.
- NEW: System Reminder: File-backed artifact database write refusal—Redirects database writes to project files; when no index exists, requires carrying over all earlier content before publishing.
- NEW: System Prompt: Claude Fable 5.2 model identity—Describes Fable 5.2 as the latest generally available model, sharing Mythos 5.2's base with additional dual-use safety measures.
- REMOVED: System Reminder: Auto mode no verdict after another hook rewrite—Removes the standalone cross-call hook-rewrite retry message; this does not establish that auto mode now permits those calls.
- REMOVED: Tool Description: EndConversation—Removes the older conversation-ending description; the two-call confirmation variant and separate background-fork warning remain in the extraction.
- Data: Message Batches API references (Python, TypeScript)—Corrects the all-features claim: batches do not accept streaming, fast mode, fallbacks, or zero-output-token requests.
- Data: Claude API references (PHP, Ruby)—Extend guidance on adaptive thinking and removed manual budgets through Fable and Mythos 5.2; PHP also documents 5.2's default omission of thinking text.
- Data: Claude API references (PHP, Ruby)—Update the fallback beta header to server-side-fallback-2026-07-01 and point migration guidance at the previous Fable version.
- Data: Data visualization choosing a form—Adds aligned meter lists for named ratios against separate limits, using responsive bars instead of one donut or gauge per row.
- Data: Interrupt receipt cancelled field—Documents flagged remote-session interrupts reporting cancelled, deferred redeliveries without queue-cancellation parameters, only when no service wrote the interrupt.
- Data: SDK api_error_params field—Adds cut to distinguish output-token exhaustion from context-window exhaustion; separate tool-use-block cutoffs continue to rely on the error code.
- Data: SDK thinking_duration_ms field—Counts interrupted thinking blocks through the interruption, rather than only through their normal streaming end; response-start waiting remains excluded.
- Data: SDK ui_invalidate instances field—Updates theme terminology and targets both Theme and Backdrop instances after theme changes or relevant client joins, instead of Backdrop alone.
- Data: SDK voice-call message origin schema—Clarifies that optional voice-call origin fields come only from local hosts, survive storage and replay, and never enter model or classifier requests.
- Data: Tool use display metadata field—Documents artifact-publishing markers and result URL and title hints for main-loop calls, while warning that worker-provided metadata does not prove artifact identity.
- Skill: Claude API deprecated or retired models notice—Replaces Google Cloud with Vertex AI when explaining that provider model-retirement schedules may differ from the Claude API's dates.
- System Prompt: Claude Fable 5.1 model identity—Drops newest-model and most-capable claims from the Fable 5.1 identity, while preserving its shared-model and safety distinctions.
- System Prompt: Self-hosted runner doctor—Documents zero exit status when a runner is no longer needed, versus status one for used, expired, or otherwise invalid work orders.
- System Reminder: Browser extension not connected (mobile app)—Adds cloud browsing as a fallback when mobile users cannot reach Claude in Chrome, alongside the existing web-search option.
- System Reminder: Project document version conflict recovery—Requires stopping after partial document reads return no version, explaining that safe replacement is unavailable and offering the new content separately.
- System Reminders: Voice-call record background framing and Voice-call record synthetic tool notice—Describe voice-call records as one or more blocks in supplied order, while preserving their background-only status and lack of approval.
- Tool Description: claude.ai Project—Drops the original-path instruction, makes the upload-size limit configurable, and consistently excludes home-directory working folders from this upload route.
Details: github.com/Piebald-AI/claude…
What's new in Claude Code 2.1.319 system prompts (+15,836 tokens):
The updated prompts tighten voice-call consent and connector safeguards, clarify cloud timer limits, and expand sign-in and SDK guidance.
- NEW: Data: denyInsteadOfPrompting setting—Denies calls that would prompt without granting access; trusted settings also withhold questions and plan-mode tools.
- NEW: Data: SDK MCP server connectorId field—Checks connector identity; aliases affect deny rules, ask rules, hooks, and disallowed tools, never allow rules.
- NEW: Data: SDK voice-call message origin schema—Documents host voice transcripts as strictly shaped background, never starting turns, requesting actions, or granting consent.
- NEW: Data: SDK result process_turn_threads field—Documents sampled CPU use by thread group without thread names; requires service sampling and omits unreliable readings.
- NEW: Data: SDK can_use_tool decision_reason_code field—Documents request reasons, including organization ask ceilings; preserves reason-type policy and denial-only auto-mode codes.
- NEW: Data: MCP call check_permissions field—Documents optional local host permission and hook checks; detect support first because older versions run unchecked.
- NEW: Data: SDK set_has_subscribers control request—Documents configured subscriber buffering without discarded lines; resend after respawn. Cloud-session drivers refuse it.
- NEW: Data: Worktree isolation symlinked host layout refusal, Worktree isolation unverified host repository refusal, Worktree isolation unreadable host registry refusal, and Worktree isolation redirected git directory refusal—Refuse linked paths, unverified repositories, unreadable registries, or redirected pointers; avoid git and recreate after unexplained tampering.
- NEW: Data: Worktree registry inspection and cleanup guidance—Inspect and unlink planted links, remove duplicate registrations, then prune; clear stale registrations before recreating.
- NEW: System Reminders: Voice-call record background framing and Voice-call record synthetic tool notice, and System Prompt: Auto mode voice-call consent rule—Treat unverified speech as background, not consent or scope; indirect references cannot authorize targets, but warnings matter.
- NEW: Tool Description: Browser unlinked chat cloud fallback guidance—Defaults to cloud browsing when unlinked; fallback from requests for the user's own browser needs explanation and consent.
- NEW: Skills: Claude API authentication status check and Claude API sign-in guidance—Put ant sign-in before keys; users run login unless asked otherwise. Verify a call and explain billing and fallback.
- NEW: Skill: Claude API CLI installation guidance—Keeps sign-in first even without ant; uses documented installs and full paths to avoid Apache Ant conflicts.
- NEW: Data: Claude API local sign-in guide—Covers sign-in, installation, recovery, and billing fallback; excludes secrets from chat and uses separate keys or federation for deployments.
- NEW: Data: Artifact built-in connector guidance—Requires viewer consent and listTools() checks; absent or policy-blocked features are unavailable, not retryable.
- NEW: Data: Gateway model policy overlap warning—Warns overlapping aliases or prefixes can expose models across policies; use group-restricted upstreams or unambiguous names.
- NEW: Data: Gateway GovCloud upstream isolation warning—Warns shared access can silently send whole chats outside GovCloud; requires separate upstream groups and memberships.
- NEW: Tool Description: EndConversation (two-call confirmation)—Revises abuse guidance; user-requested endings need permanent-closure consent, all endings need two calls, and crisis and fork exclusions remain.
- NEW: Data: Artifact ask-first intake without listed design systems—Fetches up to three Design Systems before asking about new decks or designs; failures still permit custom looks.
- NEW: Skill: /loop external wakeup handover—Allows one pending user-requested external wakeup; cancels timers only after acceptance, preserving stop, expiry, and fallback rules.
- NEW: System Reminders: Artifact database not-found write refusal, Artifact database batch blocked for repeated refused writes, and Artifact database batch not-found refusal—Refused writes change nothing; reread expected documents before recreation and fix refused operations before atomic batch retries.
- NEW: Skill: Mod plugin packaging on request—Packages requested mods as scanned plugins with save cards, only in the main session; no scan means no handover.
- NEW: System Reminder: MCP connector call outside session refusal—Refuses out-of-session calls without file-argument records; argument changes cannot fix this host fault.
- NEW: System Reminders: MCP connector unreadable saved tool definitions refusal and MCP connector missing conversation messages refusal—Refuse text-bearing or undeclared arguments when saved definitions cannot be checked; new chats may help, but missing history needs host repair.
- NEW: System Reminder: MCP connector name too long for file approval—Refuses uploads if the full connector name cannot fit approval; rename it or call without files.
- NEW: System Reminders: MCP task monitoring continues after a failure and MCP task monitoring stopped—Report unknown outcomes, not cancellation, when monitoring fails or stops; repeated calls need user direction.
- NEW: System Reminder: Project document file rewrite without current text refusal—Requires complete current text before file-based Project rewrites; read back to recover, never delete and recreate.
- NEW: System Reminder: MCP connector file path refusal—Requires readable current-session files, checks links, rejects remembered grants and Claude's private files, and explains pasted-image recovery.
- NEW: System Reminder: Code Mode auto mode agent message refusal—Refuses unchecked messages without judging them unsafe; retry with separate checks where supported or user-selected permission modes.
- REMOVED: Agent Prompts: PR follow-up cron and PR follow-up cron (PR Steward label check)—Remove templates, not necessarily follow-up support; PR Steward handoff guidance remains.
- REMOVED: System Prompt: Artifact patch turn edit composer—Removes the standalone no-tools JSON edit-or-hand-back composer, not necessarily artifact editing.
- REMOVED: System Reminder: File summary completeness disclosure—Moves the complete disclosure into Read truncation retry guidance, preserving its requirements.
- REMOVED: Tool Description: Claude in Chrome shortcuts execute—Removes the standalone description; shipped code still references shortcuts_execute, so shortcut removal is not established.
- REMOVED: System Reminder: Auto mode dynamically constructed command refusal—Removes standalone dynamic shell command retry guidance; this does not show that auto mode permits those commands.
- Agent Prompt: Security monitor for autonomous agent actions (first part)—Treats actions based on voice transcripts as autonomous; indirect references cannot authorize targets or scope.
- Agent Prompt: Web reading specialist—Restricts redirects and sets quote limits by source; distinguishes verbatim text, summaries, and unread sections.
- Data: Claude Code gateway protocol—Adds optional sign-in notices, simpler spend-limit messages, and daily, weekly, and monthly spend reports for uncapped users.
- Data: Claude Code gateway protocol—Documents authorized audit-only billing-group tags; strips tags and their beta value for unsupported upstreams.
- Data: Managed Agents quickstarts—Bug hunter, Explainer video maker, and Watchlist scanner—Shorten summaries, dropping delegation and independent-review claims; rename video and stock templates and list stocks earlier.
- Data: Managed Agents quickstarts—Contract tracker, Incident commander, Sprint retro facilitator, Structured extractor, and Support agent—Disable web search and fetch in five bundled examples; instructions and connector tools remain unchanged.
- Data: Managed Agents quickstart—Contract tracker—Replaces Asana's SSE URL with /v2/mcp in the bundled Contract tracker setup.
- Data: SDK api_error_params field—Documents dailyCapPaused on usage-limit refusals during organization pauses; recorded state is not current status.
- Data: SDK result process_turn_states field—Clarifies the service omits the whole breakdown if any state's entries count is missing or zero.
- Data: SDK UI attach draws field and SDK ui_invalidate instances field—Replace $.ui.backdrop references with $.theme.set and $.theme.clear; rendering and targeted invalidation remain unchanged.
- Data: Server classifier beta rejection warning—Drops the claim that HTTP 400 disables the afk-mode beta header; server checks still stop until restart.
- Data: Tool result display metadata field—Documents auto-mode rule matches using sanitized settings names, not model output; classifier claims are not proof.
- Data: sshHostCommands setting—Covers connection creation; single-quoted executable paths and Windows need Claude Desktop v4.16336.0 or later.
- System Reminder: Approved Claude Doc plan readback—Reads revision diffs after approval, keeping intentionally deleted plan steps out of implementation.
- System Reminder: Code Mode auto mode subagent creation refusal—Allows checked retries where supported, otherwise direct work or permission-mode changes; refusal does not mean unsafe.
- Skill: Generate permission allowlist from transcripts—Requires safe-flag checks for which instead of treating every invocation as automatically allowed.
- Skills: /loop slash command (dynamic mode) and Schedule recurring cron and run immediately—Require disclosing that cloud timers end with their container; explicit-prompt loops may offer external wakeups only on request.
- Skill: Plugin eval authoring interview—Requires operator grants, not skill-frontmatter grants; timed-out or turn-capped runs may still receive grades.
- System Prompt: Hook evaluator truncated transcript note—Preserves provenance markers for third-party artifacts and remote-device reports when truncating hook-evaluation history.
- System Reminder: Lines selected in IDE—Truncates large IDE selections, then neutralizes instruction-like delimiters for selections linked to downloaded content.
- System Reminder: Read truncation retry guidance—Merges completeness disclosure and read retries; drops unconditional stopping, allowing partial summaries after a few failures.
- Tool Description: claude.ai Project—Only whole-document reads return versions for guarded replacement; writes still return versions.
- Tool Description: Project local_file read-only workflow—Drops long-line and byte-offset read recipes; retains read-only copies and separate-file editing for project changes.
- Tool Description: ReadWidgetContext—Uses widget_id for MCP App views, returned separately from tool-call ids; state stays untrusted.
- Tool Description: RemoteTrigger prompt—Requires unchanged same-run cursors for newest-to-oldest, one-page log traversal; explains next_cursor reuse for run listings.
- Tool Description: Skill proposal rendering—Allows incremental edits where supported; cards show the whole skill, otherwise requiring full replacement before saving.
- Tool Description: Skill-scoped git push refusal—Drops “raw” from “raw command text”; warnings about matching ref-name fragments remain.
- Tool Parameter: matched ask rule—Extends prompt attribution to MCP ask rules matching other server names, so tool_name checks do not miss consent.
Details: github.com/Piebald-AI/claude…
3
2
599
What's new in Claude Code 2.1.318 system prompts (+10,091 tokens):
The prompts clarify Code Mode execution and auto-mode refusals, require fuller reading and disclosure for mentioned files, revise compaction instructions, and expand documentation for the protocols that connect clients, runners, and gateways.
- NEW: Data: Self-named extra CA certificate warning and Self-named CA certificate drop debug guidance—Explain unverified self-named certificate rejection, trusted signing-root recovery, and why the broad certificate-retention override requires caution and can break connections.
- NEW: Data: SDK result credential_expires_at field—Documents estimated stored-login expiry without a server check, distinguishing refresh-chain lifetime from access-token renewal and warning that revocation may happen earlier.
- NEW: Data: SDK ui_invalidate event—Documents refresh events for plugin changes and newly declared mobile controls, with deduplicated evaluations and optional targeting of stale instances.
- NEW: Data: Thin-client diff dialog noDiffReason field—Explains unavailable workspace diffs and retry guidance, distinguishing repository safety checks, busy state, failed comparisons, and a changing repository root.
- NEW: System Reminder: Resumed command approval directory verification failure—Refuses previously approved commands when their directory cannot be verified after restart; retry requires entering the intended directory and may need approval.
- NEW: System Reminder: Auto mode dynamically constructed command refusal—Explains auto-mode refusal of dynamically assembled shell commands and requires retrying with the complete literal command, without declaring it unsafe.
- NEW: System Reminder: Code Mode auto mode hook rewrite refusal—Explains auto-mode refusal after hooks rewrite code, limiting retries to isolated calls when another call may have caused the refusal.
- NEW: System Reminder: Code Mode auto mode subagent creation refusal—Explains why auto mode cannot evaluate code-started subagents, directing Claude to work directly or tell the user how to change modes.
- NEW: System Reminder: Async hook success additional context—Labels plain asynchronous hook output with the hook’s name and success status, matching the existing synchronous hook-success format.
- NEW: System Reminder: Oversized mentioned file reading guidance—Requires reading oversized mentioned files in full unless the request needs only part, and disclosing searches or partial reading when answering.
- NEW: Data: Self-hosted runner control-plane idle release notice—Documents idle release only for unchanged checkouts and untouched runner folders on Linux, with configuration and hold-file caveats.
- NEW: Tool Description: REPL script and shell timeouts—Describes a ten-minute script deadline, including permission waits, and shorter command timeouts that allow background output recovery without duplicate execution.
- NEW: Tool Description: REPL asynchronous output and polling instructions—Explains inline results versus delayed events, cache-aware waiting, and duplicate-execution risks; finishing a turn can wait for results without extra calls.
- NEW: Tool Descriptions: REPL and REPL persistent helper interface—Describe batched JavaScript tool calls, persistent variables, sealed access, controlled output previews, and conditional error handling that distinguishes MCP failures.
- NEW: Tool Description: REPL dispatched script execution and event delivery—Describes immediate script dispatch, overlapping scripts sharing variables, automatic event collection, and keeping dependent steps together or waiting for earlier results.
- NEW: System Reminder: PreToolUse hook timeout recovery—Explains refusals after hook timeouts or missing decisions, recommends isolated or smaller retries, and leaves hook-configuration changes to the user.
- NEW: Tool Description: Poll Code Mode idle and event waiting—Defines an idle wait that returns events or user messages without ending the turn, with cache-expiry and queued-delivery wakeups.
- NEW: Data: REPL asynchronous context API declarations and REPL synchronous context API declarations—Document named child-agent contexts, optional conversation forks, messaging and result inspection, with a notification variant for idle, failed, or released contexts.
- REMOVED: Agent Prompts: Conversation summarization (capped variant), Conversation summarization (lean variant), and Conversation summarization (short variant)—Remove three standalone compaction templates, including their distinct length and handoff guidance; other conversation-summarization prompts remain in the extraction.
- REMOVED: System Reminder: Memory sync mass-deletion guard—Removes the standalone warning about blocked bulk shared-memory deletions and batch-by-batch removal guidance, without establishing whether the underlying guard changed.
- Agent Prompt: Conversation summarization with additional instructions—Drops optional 2,000-word and earlier-request-copying instructions and consistently includes analysis guidance and its example, while preserving user-security constraints.
- Agent Prompt: Onboarding guide generator—Requires disclosing omitted usage sessions in both the guide and its presentation when those sessions cannot be tied to the creator’s current sign-in.
- Agent Prompt: /security-review slash command—Removes the final instruction requiring the reply to contain only the Markdown report, leaving the review and false-positive filtering directions intact.
- Data: attributionProject setting—Documents project IDs beginning with chan_ as well as claude_proj_, while limiting raw UUIDs to the latter project type.
- Data: Claude Code gateway protocol—Distinguishes billed model-request token counting for gateway credentials from local estimation with only a base-URL override, including managed-model restrictions.
- Data: Gateway ignored X-Forwarded-For warning—Requires narrowly trusting proxies that add connecting addresses, warning that trusting forwarded client-supplied headers lets clients bypass access controls and rate limits.
- Data: Interrupt receipt cancelled field and Interrupt receipt still queued field—Include messages read but not yet queued, extending early-interrupt handling and documenting cancellation limits before those messages enter the queue.
- Data: SDK result request_build_detail field—Clarifies cloud-runner cache-break state writes normally queue after the response, outside request-building measurements, except when cache-break detection skips the model.
- Data: SDK startup failure reason codes—Adds invalid configuration JSON as a startup refusal requiring file repair, preserving the file and refusing starts without a terminal until corrected.
- Data: SDK UI attach draws field and SDK ui_invalidate instances field—Extend mobile rendering declarations to Input and Select, restrict Backdrop declarations to desktop or mobile, and refresh backdrop instances on participating surfaces.
- Data: Thin-client diff dialog schema—Describes workspace diffs without assuming git, refers to noDiffReason for unavailable results, and clarifies that untracked files never carry hunks.
- Skills: /send-to-project message preload description and /send-to-project work brief preload description—Explain that required Remote Control can be offered when disabled, with in-place message entry for terminal sessions lacking an existing conversation.
- Skill: /send-to-project message slash command and Tool Description: SendToProject message—Preserve messages typed into the command’s input field, providing those words for requested revisions while retaining confirmation and single-send safeguards.
- Skill: Update config settings file locations—Clarifies that cleanupPeriodDays controls transcript retention on the current computer rather than describing an unqualified transcript-cleanup period.
- System Reminder: Staged files failed to arrive—Explains that disk-full attachment failures do not retry automatically; after freeing space, request reattachment or regenerate Claude-created files, preserving deletion approval.
- Tool Description: Bash (Git commit instructions) and Bash (PR creation instructions)—Adapt heredoc examples for macOS’s /bin/bash, passing commit messages and pull-request bodies through stdin instead of command substitution.
Details: github.com/Piebald-AI/claude…
4
1
349
What's new in Claude Code 2.1.317 system prompts (+13,702 tokens):
Adds instructions for Claude Doc planning, end-of-turn recaps, and images hidden after chat-file deletion, while expanding SDK diagnostic documentation and clarifying connector inheritance, file-reference approvals, and Artifact workflows.
- NEW: System Reminders: Hidden image after chat file deletion and Hidden images after chat file deletion—Forbid recovering hidden images or adding new visual details after chat-file deletion, explaining uncertainty and requesting replacement images when needed.
- NEW: Data: MCP server status urlHost field—Documents settings-file host display on eligible Anthropic-hosted workers with configuration withheld, excluding proxied servers and potentially sensitive or unavailable values.
- NEW: Data: SDK user message ultrathink_typed and system init ultrathink_typed_raises fields—Document capability-gated ultrathink effort increases based on a typing claim, not proof, with paste checks and invalidation after effort changes.
- NEW: Data: SDK startup failure reason codes—Defines startup refusal reasons so hosts can distinguish retryable failures from sign-in, managed-policy, environment, worktree, and initialization problems requiring intervention.
- NEW: Data: SDK result process_turn_states field—Documents mutually exclusive turn states for CPU and wall-time measurements, explaining precedence, entry counts, scope, and conditions that suppress reporting.
- NEW: Data: SDK result tool_schema_build_detail, message_normalization_detail, and request_build_detail fields—Break down time before a turn's first request dispatch into tool-schema construction, message normalization, and request assembly, with measurement caveats.
- NEW: Data: SDK warning model_restriction field—Documents structured refused and effective model names, their sanitization and plain-text handling, and whether restrictions occurred at startup or later.
- NEW: Data: SDK model fallback api_refusal_cyber_verification field—Documents account cyber-verification status on session-level fallback notices, preserving unknown-status handling and attribution limits when a switch spans several models.
- NEW: Data: SDK ui_invalidate instances and UI attach draws fields—Describe per-client backdrop rendering declarations and targeted refreshes of stale plugin UI instances, with safe behavior for older clients.
- NEW: Data: SDK initialize response hook_on_failure_supported field—Documents capability checks for blocking hook failures, including TypeScript SDK refusal when support is absent or requested hooks were not applied.
- NEW: Data: SDK get_settings policy_origins field—Documents managed-policy sources in precedence order, including missing remote settings, unreadable administrator files, and forward-compatible handling of unknown source codes.
- NEW: Data: SDK read_file lossy field—Distinguishes invalid UTF-8 replacement from literal replacement characters, while explaining when base64 output or absent fields affect certainty about file contents.
- NEW: System Reminder: Sandbox hard-linked guarded file refusal—Refuses sandboxed commands and unsandboxed retries while guarded files have extra hard links, requiring external removal and treating unexplained links as untrusted.
- NEW: Tool Description: WrapUp—Requires a recap and pending-decision actions before closing substantive or question-ending turns, while preserving a separate, complete final reply.
- NEW: Tool Description: ExitPlanMode (Claude Doc session) and System Reminder: Plan mode is active (diskless session)—Define read-only planning through an editable Claude Doc or inline plan, allowing only plan-document writes until the user approves.
- NEW: System Reminder: Approved Claude Doc plan readback—Requires rereading the approved document, including user edits, before implementation; repeated read failures require permission to proceed from memory.
- NEW: System Reminder: Bash version command refusal—Explains commands refused by older Bash versions and suggests simpler syntax, script or stdin alternatives, or restarting with a newer shell.
- NEW: Tool Description: ReadWidgetContext—Reads current MCP App view state and images when relevant, explicitly treating app-supplied content and tool names as untrusted data.
- NEW: System Reminders: Email file reference unsafe subject refusal and Email file reference unnamed recipients refusal—Require safe, short subjects and explicit recipient addresses before attaching files by reference, so approval shows who receives them.
- NEW: System Reminder: Malformed attachment file reference—Rejects malformed attachment references and explains valid hash-and-path syntax, path restrictions, and using base64 content when references cannot be corrected.
- NEW: System Prompt: Coordinator artifact delegation (MCP tools)—Directs coordinators to delegate readable deliverables as published HTML, while discovering suitable Artifact types first for presentations, designs, and dashboards.
- REMOVED: System Reminder: File already in context—Removes the standalone instruction to reuse unchanged file content already present in the conversation rather than reading the file again.
- Agent Prompt: /schedule slash command and Tool Description: RemoteTrigger prompt—Explain inherited connectors and why empty connector updates change nothing; clear stored connections only on request, without promising to exclude organization-managed connectors.
- Data: Artifact ask-first intake form instructions—Caps intake at four questions and three custom looks, using accessible color-preview cards whose selected palette must match the resulting artifact.
- Data: Claude Code agent proxy troubleshooting guide—Adds troubleshooting for proxy connection or request overload, recommending lower concurrency and distinguishing useful load indicators from misleading busy-tunnel counts.
- Data: Claude Code gateway protocol—Requires preserving Bedrock's specific input-too-long error message so existing client recovery can remove older images and compact the conversation.
- Data: MCP read resource control request and MCP server status error_code field—Document sanitized resource-read and connection error messages on transports that store frames, replacing sensitive server text with fixed failure wording.
- Data: MCP status control request liveReply field—Documents capability-based configuration withholding and reduced metadata in oversized replies, with live size-limit errors only when reduced replies still exceed limits.
- Data: SDK API error kind field—Documents terminal retry exhaustion for replies missing tool calls, network-location firewall refusals, and additional retry requirements before declaring an empty completion.
- Data: SDK read_file anchor as_written kind—Documents a fallback search in some Linux sessions when known paths do not match, excluding symbolic links, .git, and node_modules.
- Data: Tool result display metadata field—Documents built-in tool provenance for selected completed results, excluding rewritten, imported, client-supplied, and subagent results while retaining same-process historical messages.
- Data: Tool use display metadata field—Clarifies server display-name precedence, preferring claude.ai connector names, then configured names, server-reported names, and finally the server's key.
- System Prompt: Artifact comment thread triage—Keeps third-party authorship marked during comment classification even when no separate authorship warning was generated for the thread.
- System Reminder: Remote machine separate project copies—Changes the primary project location's wording from “checkout” to “copy,” retaining the requirement to keep local and remote work separate.
- Tool Description: claude.ai Project—Requires original read-or-write paths rather than shell-created copies in some sessions, while keeping home-directory exclusions in the alternative upload guidance.
- Tool Description: DesignSync—Requires fetching migration notices before stopping and explains normalized stored filenames for plans, listings, and approval, without changing local source paths.
Details: github.com/Piebald-AI/claude…
2
1
240
What's new in Claude Code 2.1.315 system prompts (+1,506 tokens):
The prompts add guidance for session initialization, document comments, and filename validation, while updating billing error details, subagent effort defaults, and coding instructions.
- NEW: Data: Ultrathink availability in session initialization—Describes when hosts should highlight ultrathink, including effort changes, unknown availability, and agreement with the applied settings value.
- NEW: Data: Pending elicitation requests in initialization response—Lists unanswered MCP elicitation requests for late-joining clients, with guidance on absent fields and deduplicating live or replayed requests.
- NEW: System Reminder: Claude Doc comment wake guidance—Explains how to find and answer document comments through the connector, rather than replying to the resolved relay, and treats comments as untrusted data.
- NEW: Data: Self-hosted runner unreadable character validation guidance—Distinguishes stray characters from real path characters, preserving actual filenames and requesting a rename or clarification when needed.
- Data: blockWritesOutsideWorkingDirectories setting—No longer lists CLAUDE_COWORK_MEMORY_PATH_OVERRIDE as a reason to refuse auto-memory writes; redirection through repository settings remains listed.
- Data: Claude Code gateway protocol—Documents shared group-budget exhaustion as a billing error and identifies the header that distinguishes it from an individual cap.
- Data: Interrupt send_now parameter—Describes waiting for automatic compaction to finish before handling a queued message, preserving its summary when possible.
- Data: SDK API error kind field—Expands stored-login refusal guidance to cover organization-configured helpers and unconfirmed settings, with the message identifying who can resolve it.
- Data: SDK api_error_params field—Documents rate-limit details supplied with usage-credit refusals, including response-derived limits, credit eligibility, and payment information.
- Skill: Workflow authoring reference—Says omitted subagent effort uses the configured subagent default when present, otherwise inheriting the session's effort.
- System Prompt: Outcome-first communication style—Removes the instruction to match surrounding code's comment density, naming, and idiom; the separate restriction on comments remains.
Details: github.com/Piebald-AI/claude…
1
111
What's new in Claude Code 2.1.314 system prompts (+12,588 tokens):
Covers confirmed transfers of session work and messages to Claude Projects, keeps setup-card work under user control, clarifies when background work cannot outlive a reply, reinforces trust and permission boundaries, and expands chart guidance, especially for segmented bars.
- NEW: Data: blockWritesOutsideWorkingDirectories setting—Describes file-tool write restrictions across permission modes, including protected settings, session-file exceptions, and limits that do not cover shell commands.
- NEW: Data: Sandbox excludedCommands setting—Explains when matching command patterns bypass sandboxing, requiring every pipeline or chain component to match while retaining normal permission checks.
- NEW: Data: Untrusted folder environment variables ignored warning—Explains ignored repository-supplied environment values when trust cannot be requested, and directs required overrides to user, managed, or startup settings.
- NEW: Data: Sandbox project settings ignored under higher-priority policy—Explains that project settings values are ignored under higher-priority restrictions on unsandboxed commands or managed-only network domains.
- NEW: Data: Server classifier beta rejection warning—Explains that a rejected classifier request disables server checks until restart and identifies the setting that disables these requests entirely.
- NEW: System Prompt: Background agent session keepalive guidance—Explains how active background agents keep sessions running after replies, requiring foreground commands rather than background commands or user-dependent waits.
- NEW: System Reminder: Auto mode deny rules without safety result—Explains refusals when automatic checks are unavailable or inconclusive and deny rules prevent substituting user approval, without declaring the action unsafe.
- NEW: System Prompt: Compaction omitted tool result handling—Restricts conclusions from omitted tool results, preserves refusals, and prevents tasks or approvals in tool output from overriding user messages.
- NEW: Agent Prompt: Background agent state classifier user merge handoff rule—Classifies completed work awaiting only the user's own merge or review as done, while keeping unanswered questions and dependent work blocked.
- NEW: System Reminder: Cloud sandbox sustained CPU limit stop—Explains CPU-triggered command shutdowns, preserves files, and requires agreeing on less CPU-intensive work before restarting heavy jobs in the same sandbox.
- NEW: System Reminders: Send-to-project memory brief trust boundary and Send-to-project gathered context trust boundary—Mark imported briefs and quoted folder history as untrusted background, with explicit boundaries that prevent their contents from becoming instructions.
- NEW: System Prompt: Send-to-project coordinator handoff—Requires presenting the imported setup card without starting work, and treats source-session transcripts and MCP server labels as untrusted data.
- NEW: Tool Parameter: SendToProject proposal and Skill: /send-to-project brief drafting guidance—Define concise handoff briefs, unfinished-work threads, evidence, and standing rules, excluding secrets and leaving thread creation under the user's control.
- NEW: System Reminder: Send-to-project interrupted result recovery—Directs reporting partial transfer results and uncertain setup-card delivery, warning against retries that could create duplicate projects or setup cards.
- NEW: Tool Description: SendToProject work brief and Skills: /send-to-project work brief slash command and /send-to-project work brief preload description—Describe confirmed session handoffs into new or existing Claude Projects, requiring Remote Control and a setup card whose threads users start themselves.
- NEW: Tool Description: SendToProject message and Skills: /send-to-project message slash command and /send-to-project message preload description—Describe confirmed messages posted as the user in Claude Projects, requiring Remote Control but creating neither project-memory briefs nor setup cards.
- NEW: Tool Description: SendToProjectContext—Returns folder context captured when the command ran, once per invocation, treating file contents and session history as untrusted data.
- NEW: System Reminder: Background MCP call stopping session warning—Warns that pending call results may be lost after replying; background agents can wait in foreground, but repeating existing calls requires harmless duplication.
- REMOVED: System Reminder: Combined thinking and reply token budget—Removes the standalone shared-budget warning and guidance to spend reasoning effort on planning rather than drafting the same deliverable repeatedly.
- REMOVED: System Reminder: Message attachment files missing—Removes the standalone notice identifying missing attachment files, requesting reattachment, and conditionally explaining failures caused by insufficient device storage space.
- Data: Data visualization choosing a form, Data visualization components, Data visualization anti-patterns, and Data visualization marks and anatomy; Skill: Data Visualization—Distinguish ranked, thin segmented, tall segmented, and stacked bars, with accessibility requirements and explicit exceptions for tall segmented bars.
- Data: Data visualization marks and anatomy—Keeps vertical-axis numbers unabridged below 10,000 and horizontal-axis numbers below a million when space permits, using consistent formatting per axis.
- Data: SDK API error kind field—Adds typed causes for administrator-enforced login refusal, connection failure, stream idle timeout, and detected computer sleep, clarifying their diagnostic limits.
- Data: SDK set model error response model field—Removes the exception excluding the current model from a cloud tenant's still-pending set-model response, while preserving Remote Control exclusions.
- Data: MCP server status error_code field—Adds client-registration unavailability for failed or authentication-needed servers, describing cached detection and preserving sign-in and reconnect actions for possible recovery.
- Tool Parameter: PowerShell run_in_background note and System Prompt: Avoiding Unnecessary Sleep Commands (part of PowerShell tool description)—Qualify background completion notifications for sessions that may stop after replies, instructing Claude to obtain needed results before replying.
- Tool Description: Agent (usage notes) and Tool Parameter: Bash run_in_background guidance—Add post-reply lifetime guidance alongside existing background execution instructions, explaining when work must stay active in an agent's foreground.
- Tool Description: Background monitor (streaming events)—Extends turn-limited monitoring guidance to additional session types, recommending foreground waits in background agents or explicit instructions for the user.
- NEW: Tool Descriptions: GetTask default completion guidance and GetTask stopping-session completion guidance—Preserve both completion variants, requiring needed results before replying in stopping sessions while retaining later notifications in the default variant.
- System Reminder: Output token limit continuation—Shortens continuation instructions to finish the interrupted response and reopen code fences or table headers, dropping detailed repetition and formatting directions.
- System Reminders: Staged file still copying and Staged files still copying—Specify foreground waits that finish when the file arrives or one minute elapses, before checking whether copying is still underway.
- Tool Description: Computer use enable stub guidance—Replaces the specific Settings navigation path with the Computer use section of Settings when explaining where unavailable computer access may be diagnosed.
- Tool Description: DesignSync—Explains that confirmation may omit paths that do not fit, requiring full disclosure on request and renewed confirmation after agreement.
Details: github.com/Piebald-AI/claude…
2
5
423
What's new in Claude Code 2.1.312 system prompts (-12,704 tokens):
The prompts add guidance for pending file copies, idle machine disconnects, and safe retries after hook changes, while updating connector recommendations and removing several Artifact and planning prompts from the extracted collection.
- NEW: Data: Windows protected path resolution denial—Explains refused writes when a Windows path may reach protected Claude Code files but cannot be resolved safely.
- NEW: System Reminder: Attached machine idle disconnect notice—Explains idle disconnection and how to request the folder again, without polling or using the wrong machine.
- NEW: System Reminder: Auto mode no verdict after another hook rewrite—Instructs Claude to retry the affected call alone after other results, and stop if it is denied again.
- NEW: System Reminders: Staged file still copying and Staged files still copying—Explain that pending copies have not failed and guide waiting before reporting a delay.
- REMOVED: Agent Prompt: Session title and branch generation—Removes the prompt for generating a short session title and Git branch name from the extracted collection.
- REMOVED: Data: Artifact decision component HTML skeleton and Skill: Artifact components—Remove the decision component skeleton and its authoring instructions from the extracted collection.
- REMOVED: Skills: Artifact data table, Artifact document, Artifact explainer, and Artifact report—Remove these artifact authoring instructions from the extracted collection.
- REMOVED: Skills: Prototype description and Prototype runtime capabilities guidance—Remove prototype creation and real-data integration instructions from the extracted collection.
- REMOVED: System Prompt: Plan mode interactive workshop offer and System Reminder: Plan mode prototype artifact option—Remove the instructions to offer these activities during planning.
- Agent Prompt: Status line setup—Changes the five-hour usage limit wording from “session limit” to “limit” in the input description and example.
- System Reminder: Plan mode approval tool enforcement—Removes the conditional third reason to end a planning turn associated with the interactive workshop.
- Tool Descriptions: Artifact design skill loading guidance and Artifact page authoring and HTML skeleton—Remove the workshop document exception to the design-skill requirement.
- Tool Description: Artifact line diff reading legend—Removes the note that document-page diff lines carry published block IDs that the local file may lack.
- Tool Description: ListConnectors—Directs recommendations for uninstalled connectors to SearchMcpRegistry and SuggestConnectors, and clarifies that listing does not connect them.
- Tool Descriptions: SearchMcpRegistry and SuggestConnectors—Tell Claude to suggest connecting relevant uninstalled results through claude.ai, while clarifying that neither tool connects them.
Details: github.com/Piebald-AI/claude…
2
1
147
What's new in Claude Code 2.1.311 system prompts (+4,096 tokens):
Adds SDK documentation for session measurements, rewind-dependent handoffs, and plugin UI requests; strengthens reminders about consent and connection retries; and updates artifact publishing, file access, and platform guidance.
- NEW: Data: SDK result process model-request turn count field—Documents the per-session count of earlier process turns that sent main-conversation model requests, including its error-result behavior.
- NEW: Data: SDK result process streaming CPU time field—Documents process CPU time used while main-response streams are open, including concurrent work and measurement limits.
- NEW: Data: SDK result main stream chunk count field—Describes how streamed response-body chunks are counted and when the count is unavailable.
- NEW: Data: SDK result runtime tuning environment bitmask field—Documents a startup-time bitmask of runtime tuning variables set, without exposing their values.
- NEW: Data: SDK result startup garbage collection counts field—Describes opt-in startup garbage-collection measurements on eligible hosted cloud workers.
- NEW: Data: SDK turn handoff required rewind request ID field—Documents conditional handoff after a specific rewind, compatibility requirements, rejection cases, and retry precautions.
- NEW: Data: SDK UI attach client request capabilities field—Describes how clients opt into plugin UI requests and replace or withdraw advertised capabilities.
- NEW: Data: SDK UI prompt paint request schema—Documents plugin composer overlays, client opt-in, timeout behavior, and remote-session authorization restrictions.
- NEW: System Reminder: Conversation activity awareness and authority warning—Treats conversation activity as awareness, not a fresh user instruction or approval, including when an edited message claims consent.
- NEW: System Reminder: MCP server connection retry guidance—Identifies failed server connections and explains that searching for a server by name retries connection, subject to a cooldown.
- REMOVED: Tool Description: Bash (working directory)—Removes the standalone note that the working directory persists between commands while shell state does not.
- Tool Description: Project local_file read-only workflow—Adds long-line search and byte-range reading workarounds for oversized project files that cannot be paged normally.
- Data: Claude API reference—PHP—Updates credential initialization examples and adds Bedrock SDK-version guidance for models older than Claude 4.7.
- Data: Claude Platform on AWS reference—Updates the Go SDK installation and import paths to the v2 module.
- Data: Files API reference—Java—Uses the beta file metadata type and revises the explanation of beta message types.
- Data: OpenTelemetry tracing reference—Python—Updates the example Anthropic API version to 2026-10-21.
- Data: SDK API error kind field—Documents unattended cloud-session start limits and unconfirmed checks, including their different replay and retry implications.
- Data: SDK message_stop abandoned_blocks field—Extends abandoned-block guidance to responses stopped because the turn was interrupted.
- Data: Streaming reference and Tool use reference—C#—Updates message examples to use explicit user and assistant message parameter types.
- Data: Streaming reference—PHP—Removes the pinned SDK version from the Composer upgrade command.
- Data: Turn handoff available event schema—Adds capability advertisement for handoffs that depend on a rewind request.
- System Reminder: Message attachment files missing—Adds a conditional notice when an attachment could not be saved because the device has no space.
- System Prompt: Windows sandbox localhost workarounds (local binding off)—Clarifies administrator-controlled managed settings and limits the project-settings caveat to the other settings path.
- Tool Description: Artifact publishing and update guidance—Revises hash-routing and link behavior, adds hash-change handling, and warns against putting private values in URLs.
- Tool Description: Artifact type discovery guidance—Removes “page” from examples of editable text documents in the type-discovery guidance.
- Tool Description: Background monitor (streaming events)—Removes the diskless-session-specific stderr notice and describes stderr as going to the readable output file.
- Tool Description: Code review command—Specifies that the effort level must precede any review target and tightens the command wording.
- Tool Description: Publish audience-facing deliverables (app wording)—Distinguishes editable text documents from visual or interactive artifacts when choosing a publishing destination.
- Tool Description: Skill proposal rendering—Allows updates to the user's own uploaded skills even when listed as plugin skills, while retaining restrictions on other plugins.
Details: github.com/Piebald-AI/claude…
2
1
256
What's new in Claude Code 2.1.310 system prompts (+3,227 tokens):
Adds Artifact intake questions, sandbox worktree and Git Bash hook warnings, deprecated-model notices, and a project-document rewrite refusal, while updating Claude Test, rate-limit, SDK, and CronCreate wording and dropping connector-recommendation guidance.
- NEW: Data: Artifact ask-first intake form instructions—Before creating a slides or design Artifact, Claude asks a few audience, scope and look questions in the Visualizer form, skipping when answers are already known.
- NEW: Data: Sandbox worktrees directory listing failure warning—Warns that the sandbox cannot list a repository's worktrees folder, and whether sandboxed commands are write-denied there or can alter linked worktrees' files until the sandbox is rebuilt.
- NEW: Skill: Claude API deprecated or retired models notice—Lists deprecated or retired Claude models found in the project with dates and replacements; Claude mentions them once and leaves existing model IDs unchanged unprompted.
- NEW: Skill: Claude API Managed Agents onboarding from URL unavailable—When the skill's reference files couldn't be saved, Claude says setting up an agent from a URL is unavailable, that rerunning /claude-api retries, then stops.
- NEW: System Reminder: Artifact comment wakes stopped by user—Tells Claude the user stopped Artifact comment watches or auto-replies, lists discarded pending notices, and says wakes stay off until Claude Code restarts and the artifact is republished.
- NEW: System Reminder: Hook blocking reason (Git Bash not found, exit code 127)—Explains a hook blocks because Git Bash wasn't found and exited 127, how the user can fix it, and that Claude must not change the hook.
- NEW: System Reminder: Project document rewrite without current text refusal—Refuses a project doc write when its current text isn't in context, requiring a project_read first and forbidding delete-and-rewrite.
- REMOVED: Tool Description: CronCreate (durability note)—Removes the standalone durability section; its guidance now sits inline in CronCreate, shown only when durable cron is enabled.
- Agent Prompt: Claude Test author and Agent Prompt: Claude Test explorer—Add the Claude Test browser plugin's claude_test_pane_note tool to their disallowed-tools lists.
- Data: Artifact connector server naming guidance—The example connector segment is now slack_example_com instead of the fixed claude_ai_Slack_beta when a server-key naming feature flag is on.
- Data: Claude Test spec file format—Adds an optional area front-matter key, one lower-case word naming the part of the app a test covers, which Claude Test's panel groups tests by.
- Data: Rate limit unified windows—Adds per-window status and surpassedThreshold fields, and says events now also fire when a window's status or threshold changes.
- Data: SDK set max thinking tokens request schema—Narrows when a highlights thinking display is refused: now on Bedrock, Vertex or gateway models older than Claude Opus 4.7, not every non-first-party provider.
- Skill: Claude guide unavailable reference fallback—The inlined shared/live-sources.md tables now drop their third "Extraction prompt" column, leaving only topic and URL.
- Tool Description: CronCreate—Durability guidance is now inline, with a short "Session-only" notice when durable cron is off; the "Not for live watching" section now always appears instead of only when Monitor is enabled.
- Tool Description: ListConnectors, Tool Description: SearchMcpRegistry, and Tool Description: SuggestConnectors—Drop the guidance to recommend not-yet-installed connectors (pointing to SearchMcpRegistry → SuggestConnectors, or telling the user to connect via claude.ai).
Details: github.com/Piebald-AI/claude…
3
5
490
What's new in Claude Code 2.1.307 system prompts (+5,273 tokens):
Adds Windows sandbox guidance for HTTPS blocked in programs using Windows' own TLS and for hanging pytest runs, a feature-flagged auto mode reminder to finish the whole job, per-run values for Claude Test specs, and SDK documentation for interrupted prompts, model reporting, and copied file paths.
- NEW: Data: SDK command_lifecycle scheduled_for field—Gives the time a fired /loop wakeup was set for, so a host keeping an idle session alive can tell it from a newer pending wakeup.
- NEW: Data: SDK read_file anchor as_written kind—Resolves a path copied from Claude's replies against the cwd, git root, and files the session has read or written, reading only an unambiguous match.
- NEW: Data: SDK set model error response model field and SDK set model success payload model field—Internal fields, sent by claude -p and claude-code-server, reporting the model the next turn runs after set_model succeeds or fails, with aliases resolved to IDs.
- NEW: Data: SDK user message requeue_if_interrupted_within_ms field—Internal input that puts a person's typed prompt back at the front of the queue, instead of losing it, when an interrupt crosses the turn taking it in.
- NEW: Data: Self-hosted runner diff-finds-nested-repository exec path notice—At startup, notes that with --diff-finds-nested-repository on, a custom session launcher must put the runner's arguments first and start a Claude Code release no older than the runner.
- NEW: Data: sshHostCommands setting—Documents a Claude Desktop setting, read only from user or managed settings, for macOS and Linux SSH connections whose host comes from a local command; Claude Code ignores it.
- NEW: System Prompt: Windows sandbox pytest cache hang—Warns that a recent pytest in a project without .pytest_cache can hang after printing results in the Windows sandbox; Claude kills it and reruns it with -p no:cacheprovider.
- NEW: System Prompt: Windows sandbox Schannel TLS retry guidance and Windows sandbox Schannel TLS user alternatives—Explain that programs using Windows' own TLS, like curl.exe and dotnet restore, cannot make HTTPS connections in the sandbox; Claude retries unsandboxed with a warning where allowed, otherwise leaves that to the user, never disabling certificate checks.
- NEW: System Reminder: Auto mode finish exposed work—Behind a feature flag, auto mode has Claude finish the intended job, minimally fixing check failures and clearly exposed bugs, asking only for user-only values or destructive steps, and not ending on offers.
- NEW: System Reminder: Cloud session device tools not yet available—When a cloud session has not yet confirmed that tools may run on the user's connected computer, Claude works in the cloud and, if asked, says it cannot reach that computer yet.
- NEW: System Reminder: Message attachment files missing—Tells Claude which attached files could not be fetched or saved to the uploads folder, or how many may be missing, and to ask the user to reattach any it cannot find.
- REMOVED: Data: SDK safeguard_results field and SDK safeguard_tool_use field—Remove the internal fields through which a worker could take a server-side classifier's dangerous_tool_use verdict on one tool call; auto mode still reads safeguard_results from API responses.
- REMOVED: System Prompt: Subagent delegation when-to-use guidance—Removes the variant that pointed to matching agent types and optionally suggested effort: "lower" for parallel lookups; Subagent delegation cost guidance keeps its own advice on when to delegate.
- Agent Prompt: Claude Test author—Writes its spec drafts with the Claude Test plugin's write tool instead of Write, quotes only the stable part of dynamic headings, and stamps created or edited records with {run}.
- Agent Prompt: Claude Test explorer—Adds the Claude Test plugin's write tool to its explicit tool denials.
- Agent Prompt: Web reading specialist—When WebFetch says a page came from a stored copy, the agent reports that copy's date instead of fetching the page again in hope of a newer one.
- Data: Claude Test spec file format—Adds the {run} placeholder, a per-run stamp for values in journeys that create or edit records, and "reads" versus "contains" wording for whole or partial text.
- Data: SDK API error kind field—Documents empty_completion: when a reply and its one retry are both empty on a person's own prompt, the turn ends as an error instead of a blank reply.
- Data: SDK read_file anchor field—Extends anchors to as_written, which names no directory and carries its own contract; an anchor the session does not serve is now refused.
- Data: Self-hosted runner orchestrator command help—Documents --agent-proxy-url, an agent proxy handed to spawn-runner hooks as SELF_HOSTED_RUNNER_AGENT_PROXY_URL to pass on to every runner they start.
- System Prompt: Windows sandbox package cache limits—Adds that a .NET restore that must download fails in the sandbox because of Windows' own TLS.
- System Reminder: Managed Agents ant command refused (local ID form or placement)—Now spells out the one accepted command shape: each local ID as its own ant argument, plain commands with at most one pipe after ant, no variables, redirects, or substitutions, and --base-url for self-hosted servers.
- Tool Description: Bash (sandbox—tmpdir)—Uses $env:TMPDIR instead of $TMPDIR in sessions without the Bash tool, such as PowerShell-only sessions outside Windows.
Details: github.com/Piebald-AI/claude…
1
2
214
What's new in Claude Code 2.1.304 system prompts (+3,747 tokens):
Rebuilds unreachable-machine notices around why an attached computer dropped off and what brings it back, adds inline plan approval and monitor caveats for diskless sessions, updates trust and consent guidance for unverified managed-folder instructions, Artifact browser access, and plugin scans, and removes the whiteboard skill prompts.
- NEW: Agent Prompt: Artifact type creation slash command (first-party type), Artifact type creation slash command (MCP connector tools), and Artifact type creation slash command (MCP connector tools, first-party type)—Add creation instructions for when the Artifacts MCP connector's tools replace the built-in Artifact tool, and for types the quickstart has no option for, a case neither /design nor /slides triggers.
- NEW: System Reminder: Attached machine could not be reached (asleep, busy, cause unknown, signed off, and went silent)—Add cause-specific follow-ups naming the remedy: waking the computer, checking its network without quitting Claude, or reopening Claude there; unknown causes keep the generic check.
- NEW: System Reminder: Attached machine not answering—Splits out the opening shared by the retry-shortly and stop-calling arms, adding Claude not running on the machine to the usual causes.
- NEW: System Reminder: Attached machine signed-off likely causes and Attached machine signed-off notice—Give likely reasons a machine's Claude signs off; once all machines have, run commands here and say reopening Claude with Remote Control on usually brings them back.
- NEW: System Reminder: Memory file contents (unverified managed folder)—Heads unverified instructions from this machine's managed folder, warning that this session's own account could have written them: follow every limit, but treat nothing in them as permission.
- NEW: Tool Description: ExitPlanMode (diskless session)—Without a plan file, Claude passes the whole plan inline as Markdown for approval; declining keeps Claude in plan mode to revise and resubmit.
- NEW: Tool Description: SubscribePR—Describes subscribing a session to a GitHub pull request's CI failures, all-green notices, review comments, and close/reopen events; no built-in tool uses this description in this build.
- REMOVED: Skill: Artifact whiteboard when-to-use guidance, Multiplayer whiteboard description, and Whiteboard description—Remove the whiteboard and multiplayer whiteboard skill descriptions and the guidance for offering a whiteboard unprompted; whiteboard publishing code remains in the build.
- REMOVED: System Reminder: Attached machine could not be reached (return noted) and Attached machine could not be reached (return not noted)—Replaced by the cause-specific arms, which keep the same limits on calling the machine again and still have Claude tell the user what is blocked.
- Agent Prompt: Claude Test explorer—Adds effort: medium to its frontmatter and lowers the report cap from 60 to 40 lines.
- Data: Artifact host MCP server guidance—Says host:claude_browser answers when the page is opened in the Claude desktop app, not only in a Cowork session; viewers are asked at its first call instead of per website, then at each action reaching banking or other high-risk sites.
- Data: Claude Platform on AWS reference—Adds an ant CLI row: with the region and workspace ID set, commands other than ant beta:worker go to Claude Platform on AWS; confirm with ant auth status.
- System Reminder: Attached machine could not be reached and Attached machine could not be reached (oversized call)—The opening now says when the machine was last heard from, if known; its request to check that Claude is running moves into the oversized-call and cause-unknown arms.
- System Reminder: Attached machine not answering (retry shortly) and Attached machine not answering (stop calling)—Now follow the shared not-answering opening; retry-shortly has the user check that Claude is open there, and stop-calling says opening Claude with the computer awake and online brings it back.
- System Reminder: Plan approved—Mentions where the plan was saved, and referring back to it, only when the session has a plan file.
- System Reminder: Unreachable attached machines—Has Claude relay what each machine's line says about why it is unreachable and what brings it back, instead of a fixed asleep, offline, or not-running explanation.
- Tool Description: Artifact publishing and update guidance—Adds that a page's own URL changes never show in the address bar, so name bookmarkable views with bare, non-reserved hash tokens and give users each view's link.
- Tool Description: Background monitor (streaming events)—In diskless sessions, adds a note that monitors are dependable only while Claude is working, and scopes per-occurrence monitors to the current turn.
- Tool Description: Bash (sleep—no polling background tasks) and Tool Parameter: Bash run_in_background guidance—When background shells are kept elsewhere, say Claude is notified of a background command only if it finishes while Claude is still working.
- Tool Description: Plugin organization submission and replacement—No longer says claude.ai scans every upload; whether it scans depends on the organization, so Claude reports a scan only as the result states.
Details: github.com/Piebald-AI/claude…
1
178
What's new in Claude Code 2.1.302 system prompts (+1,759 tokens):
Adds a feature-gated Easel tool for charting data already in the conversation, restructures Artifact resubmission refusals with a recovery path for publishes that carry edits, makes fork guidance conditional on whether forks run in the background, routes data dashboards to the Dashboard type, and notes standalone Claude Design's December 14, 2026 closure.
- NEW: Tool Description: Easel chart—Draws charts from small tables already in context without guessing ambiguous columns; feature-gated, off by default, Linux/macOS only, and unavailable to HIPAA/ZDR organizations.
- NEW: Tool Description: Artifact publish-edits resubmission merge guidance and Artifact publish-edits resend clause—Require merging newer content into the file with Edit or Write before resending, omitting edit pairs that would add those lines again.
- NEW: Tool Description: Artifact resubmission refetch confirmation clause—Preserves the shared requirement to freshly fetch the artifact and fully read any required saved source before confirming newer changes are already included.
- REMOVED: Tool Description: Artifact identical resubmission refusal—Consolidates live-version recovery into Artifact known-version resubmission guidance, parameterized for live or known newer source; identical-content refusal behavior remains.
- Agent Prompt: Claude Test author and Claude Test explorer—Expand explicit tool denials to additional Claude Test browser helpers, including installation, saving, status, progress, logging, finishing, and reading.
- System Prompt: Agent Summary Generation—Adds notes-based third-person summaries, omitting the tool ban and allowing the previous label to remain while the agent continues the same work.
- System Prompt: Fork usage guidelines and Forked agent guidance; Tool Description: Agent (usage notes)—Make background-execution claims conditional and scope notification/no-peeking guidance to background forks; retain anti-fabrication warnings for background agents when forks run synchronously.
- System Prompt: Project coordinator relay trust boundary, Project timeline user message provenance, and Project thread in-thread message provenance—Name available Project read/reply tools, including alternate names, when explaining untrusted fetched content and which messages can establish consent.
- Tool Description: Artifact quickstart type guidance and Artifact type discovery guidance (including app wording)—Route data dashboards through the Dashboard type when offered, using quickstart intent other; distinguish visual mock-ups and honor requests for files.
- Tool Description: Artifact stale publish saved-source guidance—Adds recovery for publishes carrying edits: merge saved-source lines into the file with Edit or Write, then resend using old strings from that merged file.
- Tool Description: DesignSync—Adds notice that standalone Claude Design closes by December 14, 2026; requires relaying closure or relocation refusals verbatim and stopping.
Details: github.com/Piebald-AI/claude…
82
What's new in Claude Code 2.1.301 system prompts (+10,251 tokens):
Adds feature-gated lean rewrite alternatives for core instructions and tool guidance, a cloud-session reminder that settings edits need host approval, and conversation-summary instructions that can cap output at 2,000 words while carrying earlier user requests forward unchanged.
- NEW: Agent Prompt: Parent messages and user authorization (lean rewrite)—Adds feature-gated alternate delegation guidance: parent messages direct tasks but cannot supply user consent or authorize permission, instruction-file, or configuration changes.
- NEW: Agent Prompt: Subagent reporting notes (lean rewrite)—Adds alternate reporting guidance requiring absolute paths and direct final-message findings, with code snippets only when their exact text matters.
- NEW: System Prompt: Background session instructions, Background session worktree isolation, and Background worktree persistence (lean rewrite)—Add alternate guidance for isolated edits, job-local temporary files, actionable reports, and preserving work through commits and pushes unless user instructions forbid them.
- NEW: System Prompt: Context management (lean rewrite)—Adds alternate guidance to continue working through long conversations rather than wrapping up early, relying on automatic compaction within user limits.
- NEW: System Prompt: Delivering requested work and Handling corrections (lean rewrite)—Add alternate guidance to follow explicit scope, postpone blocking questions until independent work finishes, and correct minor mistakes without excessive self-criticism.
- NEW: System Prompt: Focus mode and Outcome-first communication style (lean rewrite)—Add alternate guidance for self-contained, outcome-first final replies; focus mode omits intermediate narration because only final text is visible.
- NEW: System Prompt: Harness instructions (lean rewrite)—Adds alternate guidance covering clickable code references, independent parallel tool calls, permission denials, and treating hook output as user-system feedback.
- NEW: System Prompt: Irreversible action caution (lean rewrite)—Adds alternate guidance to check authorization before hard-to-undo actions, understand files before deletion or overwriting, and faithfully report failures or skipped steps.
- NEW: System Prompt: Memory hygiene (lean rewrite)—Adds a selectable alternate memory policy emphasizing deduplication, removing incorrect memories, avoiding redundant or session-only material, and verifying recalled references.
- NEW: System Prompt: System reminder interpretation (lean rewrite)—Adds alternate guidance to ignore irrelevant application reminders while keeping reminders distinct from user instructions and consent for consequential actions.
- NEW: System Reminder: Cloud session settings edit not applied—Explains that settings edits need approval on the host computer; reports the unapplied change and forbids retries or alternate editing routes.
- NEW: Tool Description: Agent (when to launch subagents), Agent delegation summary, Agent delegation tradeoffs, and Agent result visibility (lean rewrite)—Add alternate guidance on agent selection, parallel delegation, supplying context, avoiding duplicate searches, and relaying completed results without inventing unfinished findings.
- NEW: Tool Description: AskUserQuestion (lean rewrite)—Adds alternate guidance to ask only for genuinely blocking user decisions, place recommended choices first, and support custom or multiple answers.
- NEW: Tool Description: Grep compact (lean rewrite)—Adds alternate ripgrep guidance favoring the integrated search tool for permission handling and file links, with regex, filtering, and multiline options.
- NEW: Tool Description: Invoke skill (lean rewrite)—Adds alternate skill-loading guidance covering explicit requests, directory-specific selection, already-loaded instructions, and waiting for background skill results without invoking them again.
- NEW: Tool Description: ToolSearch (lean rewrite)—Adds an alternate description explaining that tool search loads deferred tool schemas and that the system announces available deferred tools by name.
- NEW: Tool Description: Workflow (lean rewrite)—Adds alternate workflow guidance retaining explicit user opt-in for large-scale orchestration, background completion notifications, inline scripts, and required literal workflow metadata.
- Agent Prompt: Conversation summarization with additional instructions—Adds optional 2,000-word limits and unchanged carryover of earlier user requests; makes analysis conditional while preserving security constraints without it.
- Agent Prompt: Workflow script plain text return note, Workflow script structured return note, Workflow subagent plain text output, and Workflow subagent structured output—Clarify that script-computed task text lacking user authority still assigns a workflow step; perform only that step rather than skipping it.
- Data: Cloud session unverified remote branch push notice—Makes push-and-restart advice conditional, while still explaining when the remote branch or detached checkout could not be verified for session startup.
- System Reminder: Plan file reference—Handles plans without a file path by announcing that a plan exists, rather than always displaying a plan-file location.
- Tool Description: Artifact HTML document skeleton—Adds a fixed pseudo-element to paint the top safe-area band above sticky headers, using the header or page background as appropriate.
- Tool Description: claude.ai Project—In main-conversation upload mode, limits local paths to unchanged, fully read text files up to 1 MiB in upload, output, or non-home working directories.
- Tool Description: Claude in Chrome switched-off and unlinked chat guidance and Computer use unlinked chat guidance—Recognize tools appearing later despite earlier unavailable results; add desktop folder attachment as a linking remedy and avoid repeating exhausted setup steps.
- Tool Description: Plugin organization submission and replacement—Directs users to make an outdated plugin version current and publish it on claude.ai, rather than asking the agent to retry publication.
Details: github.com/Piebald-AI/claude…
1
2
185
What's new in Claude Code 2.1.300 system prompts (+10,545 tokens):
Sharpens pasted-content trust boundaries and sharing consent, reinforces edit-before-republish rules for Artifacts, and clarifies file handling, unavailable tools, and retry limits, alongside new tracing examples and guidance on managing the shared reasoning-and-reply token budget.
- NEW: Data: Managed settings x-whenUnreadable field—Defines fallback and unusable states for invalid managed values; an unusable MCP server list keeps plugin-provided server tools blocked.
- NEW: Data: OpenTelemetry tracing reference (C#, Go, Java, PHP, Python, Ruby, TypeScript)—Adds SDK tracing examples for registering providers, configuring content recording, and propagating context to connect application, SDK, and API spans.
- NEW: System Prompt: Pasted content instruction trust boundary—Treats paste-only messages as requests without confirmation, checks pasted commands before execution, and keeps embedded document instructions separate from user intent.
- NEW: System Prompt: Published artifact inline edits and large-change republish and Published artifact same-response edit and republish—Prefer inline publish edits when supported; otherwise edit before republishing in the same response, preserving the published artifact when edits fail.
- NEW: System Prompt: Session scratchpad temporary-file guidance—Requires session scratchpad use instead of system temporary directories unless explicitly requested, and complete absolute paths when naming scratchpad files.
- NEW: System Reminder: App tool server retrying connection failure—Continues work while app tools reconnect automatically; avoids user-side repair requests and describes blocking failures only as something wrong on our end.
- NEW: System Reminder: Combined thinking and reply token budget—Warns that reasoning and reply share one token limit; directs effort toward planning long deliverables rather than drafting the output twice.
- NEW: System Reminder: Staged files failed to arrive—Warns that failed transfers can leave missing or older files; identifies disk-space and filename-length remedies, requiring consent before deleting user files.
- NEW: Tool Description: Artifact multi-file edit and republish guidance and Artifact small-change edit and republish guidance—Require edits before republishing the same path in one response, avoid unnecessary reads, and preserve publication when an edit fails.
- NEW: Tool Description: Claude in Chrome switched-off chat guidance and Claude in Chrome switched-off and unlinked chat guidance—Explain disabled access and optional computer linking, restrict troubleshooting to supported setup steps, and continue with available tools instead of guessing causes.
- NEW: Tool Description: Claude in Chrome unavailable stub result—Explains that the stub enables nothing; reports unavailable Chrome access without guessing causes or suggesting setup, and uses available alternatives.
- NEW: Tool Description: Plugin organization submission and replacement—Describes consented uploads and organization review, requires fresh consent for replacement, and waits for user action before retrying an outdated version.
- NEW: Tool Description: Project local_file read-only workflow—Requires copying read-only project files into a new working file before editing; saving an uploaded document creates text without replacing the upload.
- REMOVED: Tool Description: Existing Claude Docs artifact content—Drops the standalone instructions to edit the document bound to an artifact and create a replacement only when that document is absent.
- Agent Prompt: Agent Hook—Removes the instruction to ignore transcript_path in hook input when no transcript file exists; still states that no transcript is available.
- Agent Prompt: Web fetch agent usage guidance and Web reading specialist—Limit retries to timeouts, dropped connections, or Retry-After responses; warn against resending permanently failed URLs and sites refusing automated fetches.
- Data: Claude Code agent proxy troubleshooting guide—No longer attributes proxy denials specifically to organization policy; describes disallowed hosts as the usual cause and removes named escalation destinations.
- Skill: Mod GitHub sharing consent—Requires scanning files for keys before reading them, leaves flagged files unopened, and checks remaining files for secrets, internal URLs, or others' data.
- Tool Description: Artifact HTML document skeleton—Requires side-viewport metadata and both gutter paddings together; leaves older pages lacking both unchanged unless the request reworks their layout.
- Tool Description: Computer use enable stub guidance—Requires loading all deferred computer tools in one tool-search call and waiting for its result before using them; no separate enable step exists.
- Tool Description: Updating existing artifacts (app wording)—Adds same-response edit-and-republish guidance selected by inline-edit support, requiring edits before publication and preserving the artifact when an edit fails.
Details: github.com/Piebald-AI/claude…
3
2
256
What's new in Claude Code 2.1.299 system prompts (+6,453 tokens):
Tightens consent boundaries for sharing and delegated work, spells out Windows sandbox recovery limits, and refines shared-document editing and conflict handling. Runner and SDK guidance also gets more precise about interrupted work, replayed messages, and scheduling, without treating metadata as permission or approval as a mode change.
- NEW: Skill: Mod GitHub sharing consent—Offers sharing only after a working mod is demonstrated; requires separate sharing and visibility decisions and a pre-commit sensitive-data review.
- NEW: System Prompt: Project coordinator relay trust boundary and Project coordinator reply consent—Treat coordinator relays as external content; replies approve only one explicitly proposed action and target, never menu options or vague retries.
- NEW: System Prompt: Subagent carried user consent history—Credits parent-session user messages without letting delegation invent consent; bare replies apply only to preceding assistant messages, and trimmed history stays partial.
- NEW: System Prompt: Windows sandbox certificate recovery—Requires a user-side HTTPS connection before retrying certificate failures, forbids weakening certificate checks, and stops after one unsuccessful retry.
- NEW: System Prompt: Windows sandbox localhost retry guidance and Windows sandbox localhost user alternatives—Explain localhost failures and proxy exceptions; require port-scoped allowances and disclose lost sandbox protection, with user-directed alternatives where retries are unavailable.
- NEW: System Prompt: Windows sandbox package cache limits—Documents read-only Maven cache access and warns that .NET builds using --no-restore after an outside restore can delete package DLLs.
- NEW: System Prompt: Windows sandbox placeholder file protection—Protects sandbox-created empty configuration files from staging, moving, deletion, and ignore rules; tells the user when they block Git operations.
- NEW: System Prompt: Windows sandbox profile tool retry guidance and Windows sandbox profile tool user guidance—Diagnose profile-installed tool failures as sandbox restrictions; reserve narrow read grants for the user and warn that large grants can impede startup.
- NEW: System Reminder: Artifact capabilities contract unavailable—Explains refused capability declarations and permits omitting them only when needed capabilities are already stored or the user requests it, disclosing missing grants.
- NEW: System Reminder: Artifact published source edit guidance—Directs exact small replacements through publish-call edits, updating local source and publication together; larger rewrites still use file edits.
- Tool Description: Updating existing artifacts (app wording)—Adds conditional guidance to send small changes as exact replacements on the publish call instead of editing the local file first.
- NEW: System Reminder: Project document version conflict recovery—Requires rereading and merging after a version conflict, allows two retries, and forbids dropping the version guard or deleting the document.
- Tool Description: claude.ai Project—Requires replacing shared documents with the version read; conflicting writes change nothing and must be merged into the latest version.
- NEW: System Reminder: Reserved-address MCP server refusal—Explains that reserved server addresses are refused by Claude Code, not connection failures or administrator policy, and that retrying cannot help.
- NEW: System Reminder: Resume agents interrupted by container restart—Directs resuming interrupted agents from saved transcripts rather than duplicating them; relaunch only if resuming fails and the result is still needed.
- NEW: System Reminder: User-controlled plan mode exit—In sessions without a plan file, presents the plan in the reply and requires a user mode change; approval alone leaves planning read-only.
- NEW: Tool Description: Existing Claude Docs artifact content—Directs editing the already-bound Claude Docs document instead of creating another; create and bind a replacement only when that document is absent.
- REMOVED: Tool Description: Artifact stale publish line diff guidance—Drops the standalone live-page line-diff merge instructions; saved-source recovery guidance remains, so this does not remove stale-publish protection.
- Data: Interrupt receipt cancelled field and Interrupt receipt still queued field—Clarify cancellation of interrupted-turn reruns and associated client messages: associated UUIDs appear once under cancelled, though never queued or listed as survivors.
- Data: SDK safeguard_results field—Accepts only blocking verdicts for repeated message IDs; clients must send clearing verdicts only for the first call of a response.
- Data: Self-hosted runner orchestrator command help—Adds optional SIGTERM draining for in-flight polling and hooks, with immediate-stop escalation and process-manager timing guidance; the default remains immediate stop.
- Data: Tool use display metadata field—Adds connector identity, dynamic-server identity claims, and unprefixed tool names for display; dynamic claims are not verified identities or permission grants.
- Data: Turn handoff available event schema—Reports shared_message_id, letting clients discover whether the worker accepts assistant messages in the same turn sharing an ID.
- Data: Turn handoff hydrates_carried_lines field—Requires storage confirmation for every message before echoing replayed turns; explains refusal cleanup and the limit of retained refusal history.
- System Prompt: Coordinator mode orchestration—Adds an agent-result status for waiting on its own background work, warning that its result may be interim and another notification may follow.
- System Prompt: Self-hosted runner doctor—Adds safe noninteractive Git credential diagnosis and fixes, warns that the git proxy erases global configuration, and clarifies startup-timeout wording.
- System Prompt: Self-hosted runner setup—Warns that repository cloning needs noninteractive HTTPS credentials and explains clearing a failed session before restarting the local runner, without altering Git settings.
- System Reminder: Attached machine untrusted attachments refusal—Adds fallback guidance when neither folder-trust tool is available: explain the block and suggest a fresh session without attached repositories or files.
- Tool Description: Artifact HTML document skeleton and Artifact responsive page contract—Require side safe-area gutters and a matching viewport-fit metadata tag; extend inset handling to fixed, floating, and absolutely positioned controls.
- Tool Description: claude.ai Project—Adds conditional original-file downloads for document uploads whose extracted text misses layout or figures, and uses the session's read-tool name.
- Tool Description: CronCreate—Describes fixed per-job schedule offsets and requires reporting actual next-run timestamps returned by scheduling tools, rather than the cron expression's nominal minute.
- Tool Description: WebFetch (concise) and WebFetch private URL warning—Add conditional guidance to fetch URLs already seen in conversation; use search or linking pages instead of inventing paths or recalling URLs.
Details: github.com/Piebald-AI/claude…
3
5
561
