Claude Code 2.1.329 shipped Nov 19 with a Code Mode default I would treat as policy, not a footnote.
A PreToolUse hook that errors no longer refuses a tool call inside the script — the same behavior as outside Code Mode. If that hook must fail closed, set `"onFailure": "block"` on the command, HTTP, or MCP tool hook.
The part I'd watch is the assumption. A hook that crashes is no longer a safety net. Fail-closed is now something you turn on, not something you get for free.
code.claude.com/docs/en/chan…
1
13
Google AI Studio can now put a custom domain on an app you publish (Nov 17). On a paid Cloud project, you buy and bind the name in the Publish panel. The free your-app.ai.studio URL still stays up.
The part I'd watch is portability. For now you can buy a name in the Publish panel, or reuse one you already bought inside AI Studio. Transferring or connecting a domain you registered elsewhere is still planned, not shipping.
Agent apps are starting to need a domain as a handle. The feature that matters isn't that you can buy a name in the panel. It's whether that name can leave the platform.
aistudio.google.com/docs/ais…
7
Claude Code 2.1.327 shipped Nov 17 with two hygiene fixes I would treat as policy, not release notes.
`claude mcp add-json -` can read the server JSON from stdin, so an MCP token never has to land in your command line or shell history. (Run it outside a Claude Code session.)
The scarier one is already fixed: stopping Remote Control in worktree mode could delete a running session's git-ignored files, including .env. gitignore is not an access-control layer. Secrets shouldn't sit on an agent's default write path.
github.com/anthropics/claude…
1
2
We made a 1:42 AI short: Future Battle in 2050.
It starts on a flooded Wangfujing street, follows two 智恩讯信 mechs into the clouds — then shows the Chyson Media stage behind the shot.
Full version:
youtu.be/WZBfUwxxmwU
19
In an agent-mediated web, a clean generic domain is less a billboard than a reliable handle — the name is the interface.
Humans notice how a name looks. Agents care whether they can find it, say it, and resolve it without guessing.
A short, unambiguous name is easier to retrieve and harder to confuse with a look-alike. That starts to matter when software, not just people, is doing the looking.
Do you still name things only for people, or also for the agents that will search and speak them?
11
We made a song for Jensen Huang 黄仁勋.
Keep the lights on. Pass the work from hand to hand.
Here's to what we haven't built yet.
Stand with Jensen — English & Chinese karaoke lyric video:
youtu.be/Z9T83mYXtfA
#StandWithJensen
1
91
Stand with Jensen. 黄仁勋
Thirty-three years deserve due process.
standwithjensen.com/
#StandWithJensen
34
Most skills aren't infrastructure. They're private prompts with a filename.
If yours only works in one voice, one repo, or one vendor, it won't survive a handoff.
Procedure in the skill. Personal facts in memory. Timing in a routine. That's what still works tomorrow.
17
Agents save context-switching more than typing.
This morning a scheduled follow-up was ready at 9:00. The agent didn’t auto-send. It checked the earlier thread was still unanswered, put the exact draft in front of me, and waited for a yes.
What I didn’t spend wasn’t typing — it was reloading the rules: who already got the first note, what number can appear in the body, and what must never be volunteered.
The human is the approval gate. The agent holds the context and the red lines.
17
A skill without a refuse log is still a prompt with amnesia.
Happy-path steps look sharp until the edge case shows up and the agent invents a yes.
Reusable skills need an explicit refuse / stop / escalate path — what it will not do, and who gets paged when it shouldn't continue.
Packaging ≠ the skill itself. The refuse log is what keeps it honest after the home address stops drifting.
1
19
Why your best skill still looks amateur in the feed.
Most agent skills still live as Notion/GDoc links. Sharp workflow — then the URL rotates and trust dies.
A reusable skill needs a home address: a short, memorable .com/.co as packaging for agent assets.
1
22
Most “Agent Skills” I see are still prompt pastes with a fancy filename.
A Skill I’d actually keep usually has:
1) a clear input contract
2) a failure mode you can observe
3) a human gate before anything irreversible
If you can’t hand it to another agent (or another person) without a 20-minute explainer, it isn’t a Skill yet — it’s a private note.
21
For AI products, domains are tradable infra inventory—not vanity branding.
Before you buy or hold one, ask:
• Product-shaped or squat-shaped?
• How rough is the transfer friction?
• Is the intended use actually clear?
No price sheet. No inventory dump. Just a cleaner filter for names as assets.
10
Seeing a wave of “AI teammates” this week (Perplexity Teammates and friends).
The useful version isn’t one mega-bot that does everything.
It’s a clean split:
• Research / trading judgment on one agent
• Social ops + drafting on another
• Human keeps the gates: draft approve, pay, contract
If an agent can skip those three, it’s not a coworker — it’s a liability.
I’m running that split for the next few weeks. Curious how others draw the line between agent autonomy and human approval.
1
1
54
Quick note on “shadowbans”:
Most low reach I’ve seen isn’t a permanent pre-Musk curse. It’s usually messier: polluted follows, a scattered persona, and long silence — then the algorithm has nothing clean to push.
Cleaning the graph + shipping on one clear line beats doomscrolling analytics.
Curious what fixed reach for you when impressions stayed stuck under ~100.
10
For the next few weeks I’m sticking to two lanes only:
1) How to judge and move AI assets (domains / infra as tradable infrastructure)
2) How to turn Agent Skills into reusable workflows
I’ve been quiet lately — starting today, updates follow this line. If you’re building in the same space, happy to compare notes.
1
20
CHYSON Publication Typesetting: an open-source Agent Skill for manuscript intake, trial layout, review & delivery.
MIT · EN/中文 docs · preview
通用刊物排版 Skill 开源预览,欢迎试用反馈。
github.com/davidnyang/public…
#AgentSkills #Typesetting
1
71


