
Learn how to code faster with AI in 5 mins a day. Get high signal AI software development news, research papers & trending resources. By @superhuman_ai
Joined November 2025
- Tweets126
- Following145
- Followers5.5K
- Likes205
How to open Claude Code preview links on your phone:
Claude Code's preview links point to localhost, so they die the moment you leave your desk. A dev shared a one-paragraph fix: make the agent publish previews over Tailscale.
1. Install Tailscale on your computer and phone, same account (free).
2. Add this instruction to your project's CLAUDE.md or AGENTS.md:
Any time you give me a link to something running on this computer (a preview, a review page, a dashboard), give me a Tailscale link, never localhost. I work across several computers and my phone. Publish it with tailscale serve, open the link yourself to check it works, then send it.
3. Ask Claude Code to run the project and send a preview link. It publishes the port with tailscale serve, checks the link works, then sends an HTTPS URL that opens on your phone.
What's the first preview you'd rather open on your phone than at your desk?
144
How to run Opus 5.5 and GPT-6.1 Sol overnight without one stuck step wasting your night:
1. Install Herdr with brew install herdr, and give Claude Code its skill: npx skills add herdrdev/herdr --skill herdr -g
2. Run herdr in your project. Open claude --model claude-opus-5-5 in one pane and codex -m gpt-6.1-sol in the pane next to it.
3. Save this to CLAUDE.md: "For every step: plan, build, verify, fix. A step is done only when verify passes. Max 3 attempts per step, each with a specific change. Save progress after every pass. If a step still fails, record the error, mark it blocked and move to the next independent step."
4. Tell Opus that the Codex pane is its reviewer. After each step, Opus sends Sol the diff through Herdr, and the step isn't done until Sol signs off.
5. Type /goal in the Opus pane, followed by what "done" means tonight.
6. Press ctrl+b q to detach, keep the lid open, and go to sleep.
What do you let your agents run while you sleep?
teapawt.org/Voxyz_ai/status/212207โฆ
opus 5.5 + gpt-6.1 sol is f*cking insane for overnight work
1. install Herdr (brew install herdr) and give Claude Code its skill:
npx skills add herdrdev/herdr --skill herdr -g
2. run herdr in your project. one pane: claude --model claude-opus-5-5. the pane next to it: codex -m gpt-6.1-sol
3. save this to CLAUDE.md so ๐ผ๐ป๐ฒ ๐๐๐๐ฐ๐ธ ๐๐๐ฒ๐ฝ ๐ฐ๐ฎ๐ป'๐ ๐ฒ๐ฎ๐ ๐๐ผ๐๐ฟ ๐๐ต๐ผ๐น๐ฒ ๐ป๐ถ๐ด๐ต๐:
"For every step: plan, build, verify, fix. A step is done only when verify passes. Max 3 attempts per step, each with a specific change. Save progress after every pass. If a step still fails, record the error, mark it blocked and move to the next independent step."
4. tell Opus the Codex pane is its reviewer: after each step, it sends Sol the diff through Herdr, and ๐๐ต๐ฒ ๐๐๐ฒ๐ฝ ๐ถ๐๐ป'๐ ๐ฑ๐ผ๐ป๐ฒ ๐๐ป๐๐ถ๐น ๐ฆ๐ผ๐น ๐๐ถ๐ด๐ป๐ ๐ผ๐ณ๐ณ
5. in the Opus pane, type /goal and what "done" means tonight, then ctrl+b q to detach. keep the lid open and go to sleep
3
1
345
How to stop fixing the same design mistakes from your AI agent:
1. Install the learn-taste skill in Claude Code: npx skills@latest add emilkowalski/skills --skill learn-taste
2. Correct your agent's design choices during the session the way you normally would.
3. Run /learn-taste at the end of the session.
4. Let it turn every design correction into a strict rule with reasoning, so the next agent doesn't make the same mistake.
5. Point it at older sessions when the same fixes keep coming back: /learn-taste look at my last 3 sessions too, I keep fixing the same things
What design mistake do you fix most often?
teapawt.org/Voxyz_ai/status/212128โฆ
Highly recommend installing /learn-taste in Claude Code.
Simply put, it collects ๐ฒ๐๐ฒ๐ฟ๐ ๐ฑ๐ฒ๐๐ถ๐ด๐ป ๐ฐ๐ผ๐ฟ๐ฟ๐ฒ๐ฐ๐๐ถ๐ผ๐ป you made during a session and turns each one into ๐ฎ ๐๐๐ฟ๐ถ๐ฐ๐ ๐ฟ๐๐น๐ฒ ๐๐ถ๐๐ต ๐ฟ๐ฒ๐ฎ๐๐ผ๐ป๐ถ๐ป๐ด, so the next agent doesn't make the same mistake. A great time saver if you correct your agent often.
Install it, then run one line at the end of a session ๐
npx skills@latest add emilkowalski/skills --skill learn-taste
/learn-taste look at my last 3 sessions too, I keep fixing the same things
4
2
280
How to make your agent tell you what to build next every morning:
Garry Tan made a free skill called idea-miner that does exactly that. Here's how it works:
1. Have your agent read the last 24 to 48 hours of real activity every morning: meeting notes, new or changed pages in memory, and today's and yesterday's daily notes.
2. Give it the list of skills you already have, so it never pitches the same idea twice.
3. Ask it to find the jobs that showed up more than once: manual, repetitive, or judgment work an AI could take over.
4. Make every idea point to something real from the last 48 hours, with a one-line spec that names it and says when it runs and what it does.
5. Drop any idea that already exists as a skill, or pitch it as a specific upgrade to that skill.
6. Save each run as a dated page, like idea-miner/YYYY-MM-DD, so the next run shows which ideas moved up, moved down, got built, or went quiet.
7. Start the output with "If you build one thing this week, build this:" and keep it to 3 to 6 ranked ideas with no filler.
8. Run it as a daily job at 7am.
What would your agent tell you to build this week?
teapawt.org/garrytan/status/212135โฆ
What if your own personal agent could recursively self improve as you use it? You can. I made the idea-minder skill to do exactly that.
gbrain.io/s/idea-miner-what-โฆ
240
How to turn your team's dev process into AI agent skills:
1. List every process your team repeats, because nearly every process in the SDLC warrants a custom skill.
2. Create a /create-ticket skill, so every ticket gets written the same way.
3. Add /fix-bug and /implement-ticket, so your agent follows your team's steps for both kinds of work.
4. Add /review-code, so every review checks the same things.
5. Add /open-pr, so every pull request goes out in the same shape.
6. Add /fix-pr-comments, so your agent works through review feedback for you.
7. Improve each skill whenever it gets something wrong. A skill lets your team standardize, automate, and continually improve the process.
Which process would you turn into a skill first?
teapawt.org/housecor/status/212133โฆ
Nearly every process in the SDLC warrants a custom skill:
/create-ticket
/fix-bug
/implement-ticket
/review-code
/open-pr
/fix-pr-comments
If our team repeats a process, a skill means we can standardize, automate, and continually improve the process.
1
167
How to make Claude Code fix its own failing checks before it stops:
1. Add a Stop hook that runs your check, like your tests or your linter, so Claude can try again when the check fails.
2. Exit with code 2 when the check fails. Exit 2 sends Claude back to retry with the hook's stderr.
3. Never use exit 1 for a failed check. Exit 1 is non-blocking, so Claude just finishes even when the check failed.
4. Send your command's output to stderr, so Claude can see what went wrong.
5. Copy this pattern: [your command] >&2 || { echo failed; exit 2; }
6. Treat an exit 1 as a possibly broken hook, because it can also mean the hook itself errored, such as a missing dependency or a bug in the hook script.
What checks do you make Claude pass before it stops?
teapawt.org/lydiahallie/status/212โฆ
You can use a Stop hook to have Claude try again when a check fails, but make sure it exits with code 2!
Exit 2 sends Claude back to retry with the hook's stderr, exit 1 is non-blocking so Claude just finishes even when the check failed
1
2
186
How to stop using Fable 5.2 like an intern:
1. Hunt the hidden state machines. Paste: "Find all of the implicit state machines in my app. Make the best decision that will elevate all three: the user experience (UX), developer experience (DX), and agent experience (AX) always, without breaking anything."
2. Scout open source before you build. Paste: "Don't write an unfamiliar feature from scratch. Search GitHub for open-source projects, compare their approaches, then start."
3. Triage bugs before you fix them. Paste: "Rate each bug from 1 to 10 on how often it would happen and how bad it would be. If the two scores multiplied are below 5, just file it as an issue and leave the code alone. Do not design for rare or future cases nobody asked about." Tune the threshold of 5 to your project.
4. Make it finish the job. Paste: "Finish the whole job before you stop. Don't offer to continue: continue. Ask me before you delete anything. End with a recap for someone who saw none of the work: key points first, plain words."
5. Audit your CI before you try to speed it up. Paste: "Audit this repo's CI: Use gh to pull the last 150 CI runs and split each job's time into three parts: waiting in the queue for a machine, setup (checkout, installing dependencies, restoring the cache), and actually running the task. List the 5 slowest jobs and how many seconds each part takes. Give me the report first. Don't change anything."
6. Check the plan from scratch. Ask: "If you were starting over today with the same goal, what would you still need to keep?"
Which prompt will you try first?
teapawt.org/Voxyz_ai/status/212201โฆ
Don't tell Fable 5.2 "find the bugs."
Don't tell Fable 5.2 "build this feature."
Don't tell Fable 5.2 "why is CI so slow."
You're using the smartest coding model out there ๐น๐ถ๐ธ๐ฒ ๐ถ๐'๐ ๐ฎ๐ป ๐ถ๐ป๐๐ฒ๐ฟ๐ป.
Here are 6 prompts to copy and paste:
172
How to stop AI agents padding your UI with helper text:
AI-built screens arrive wrapped in taglines, helper text and secondary labels that explain a layout instead of fixing it. Vox shared a one-line AGENTS.md rule that fixes it.
> Step 1: Open AGENTS.md in your project root (Claude Code also reads CLAUDE.md). Create it if it doesn't exist.
> Step 2: Add this line:
- Cut taglines, helper text and secondary labels. If a screen needs them to make sense, fix the layout.
> Step 3: Start a new session.
The agent now treats explanatory copy as a layout bug and reworks spacing, hierarchy or grouping instead of adding text.
teapawt.org/Voxyz_ai/status/211325โฆ
Vibecoders, ๐'๐บ ๐ฏ๐ฒ๐ด๐ด๐ถ๐ป๐ด ๐๐ผ๐, add these two lines to your AGENTS.md:
"Cut taglines, helper text and secondary labels. If a screen needs them to make sense, fix the layout."
"Make the best decision that will elevate all three: the user experience (UX), developer experience (DX), and agent experience (AX) always, without breaking anything."
3
1
249
Found a repo that lets your AI agents ask each other for help, wherever they run.
Every agent dials out to a small relay on your Tailscale network, so nothing needs an open port and there are no API keys between agents.
Github: github.com/mvanhorn/agent-tiโฆ
1
171
If your app renders Mermaid diagrams, this repo does it with a much smaller bundle.
From Obsidian creator kepano: feed it Mermaid syntax and get back themeable SVG, with zero runtime dependencies.
198
How to run an agent over millions of rows without blowing your budget:
Andrej Karpathy built a better YouTube feed out of 43 million Hacker News comments, without dumping them all into a model.
The comments were filtered first with URL matching, regexes and spam removal. Only what was left went to Claude Sonnet. About two hours later he had 6,454 videos Hacker News users keep recommending.
> Step 1: Before any model call, ask your agent for a filter script:
Write a script that filters [data] down to only the rows that could matter for [question]. Use regexes, URL matching and dedupes. Don't call a model. Print the row count before and after.
> Step 2: Check the before-and-after count, and tighten the filter if too much is left.
> Step 3: Switch to a cheaper model, then ask Claude to read the filtered file and answer your question:
/model sonnet
The same approach works on any large log or dataset.
teapawt.org/karpathy/status/211697โฆ
Replying to @eyadnour
Oh tbc Claude filtered it down a lot: urls, regexes, dropping spam, only the remainder was dispatched to Sonnet with a CC Workflow (so it uses the sub tokens). You can see surviving comments under videos, the comments that recommend videos seem to be of nice & helpful character.
203
How to stop your coding agents from copying old hacks across your codebase:
1. Delete most of the comments in any codebase your agents work in.
2. Know why it matters: agents love copying whatever is already in the codebase.
3. Watch out for workaround comments, because an agent reads one as "workarounds are fine here."
4. Expect copied hacks and stale comments to spread across the whole codebase if you leave them in.
5. Send this to Claude Code or Codex: "Identify the useless code comments we can prune in my repo. Keep the ones that actually should exist (like explaining a browser quirk). Come back with a high level plan of which of these workarounds would be best solved by refactoring and which we can straight up delete."
6. Keep the comments that truly need to exist, like one that explains a browser quirk.
7. Review its plan, then refactor the workarounds worth fixing and delete the rest.
Highly recommend ๐ฑ๐ฒ๐น๐ฒ๐๐ถ๐ป๐ด ๐บ๐ผ๐๐ ๐ผ๐ณ ๐๐ต๐ฒ ๐ฐ๐ผ๐บ๐บ๐ฒ๐ป๐๐ in any codebase your agents work in.
Simply put, agents love copying whatever is already in the codebase. Leave one comment explaining a temporary workaround and they read it as "workarounds are fine here." Before long, the whole codebase is ๐ฐ๐ผ๐ฝ๐ถ๐ฒ๐ฑ ๐ต๐ฎ๐ฐ๐ธ๐ ๐ฎ๐ป๐ฑ ๐๐๐ฎ๐น๐ฒ ๐ฐ๐ผ๐บ๐บ๐ฒ๐ป๐๐.
Send this to Claude Code or Codex ๐
"Identify the useless code comments we can prune in my repo. Keep the ones that actually should exist (like explaining a browser quirk). Come back with a high level plan of which of these workarounds would be best solved by refactoring and which we can straight up delete."
279
How to get far more out of your coding agents by thinking like an executive:
1. Send your prompt and move on to the next problem instead of staring at the chat to watch the agent work.
2. Ask your agent for a document or presentation before high-stakes decisions instead of deciding everything in plain chat.
3. Fix the mistake when your agent gets something wrong, then find what allowed it to happen and stop it at the root.
4. Stop being the tester who checks every line for bugs, because no popular product on earth is bug-free.
5. Invest in workflows that catch bugs early before you ship.
6. Set up alerts so you know when a bug happens, and make sure fixes can go out quickly.
7. Ask yourself this before every task: "If I had 100x more agents doing 100x more work for me every day, how would I do this?"
8. Repeat that question until it becomes a habit, because the mindset lasts even as tools, skills, and workflows keep changing.
i found "executive mindset" to be the biggest differentiator from people who can use agents super effectively
it comes natural for people who had been an executive of a large org before, but it's counter-intuitive for people who haven't done so
a few examples
1. after you send a prompt, do you stare at the chat to see what the agent is doing?
most agent UX today leads people to do that, but a good executive never watches over an employee's shoulder after they've sent an ask - they are immediately on to the next problem
2. do you rely on chat for important decision making?
"just chat" sounds elegant, but there's a reason company executives rarely rely on plain talking for high stake decisions
instead, they ask the team for documents or presentations. it puts a bit more prep work on the team, but the discussions end up being a lot more efficient
3. when your agent did something wrong, do you curse at it and twist its arms to do it right?
some successful executives actually do that as well, but they also do something else afterwards - figure out what allowed the mistake to happen in the first place, and stop it at its root
4. do you spend all your time carefully examining everything to make sure the code is bug free?
now this is getting spicy, but you can pick any popular product on earth and you won't find a single one that's bug free
good executives accept the calculated risk and invest time in the right workflows to catch bugs early before shipping, get alerts if they happen, and making sure fixes can go out quickly - never make themselves the bottleneck by being the tester
our tools, skills, workflows etc change all the time, but building the right mindset is durable and can help you for a long time
if you aren't clear how to acquire the mindset yet, i recommend starting with asking yourself this when you try to decide to do something or not -
"if i have 100x more agents doing 100x more work for me every day, how would i do this?"
do this enough times, and you'll find yourself actually able to control 100x more agents effectively
1
315
How to make Claude Code and Codex fix the rough edges in your app:
1. Know why agents miss polish: they build for the demo, where everything goes right, and real use never works that way.
2. Ask your agent to find all the happy-path debt in your app.
3. Have it check every screen that only works when data loads fast, nothing fails, lists aren't empty, text is short, and the screen is wide.
4. Tell it to fix missing states and layout breaks one component at a time.
5. Make it keep your design as it is and avoid breaking anything.
6. Ask for a summary by file of how each part worked before and how it works now.
7. Paste this prompt: "Find all the happy-path debt in my app: every screen that only works when data loads fast, nothing fails, lists aren't empty, text is short, and the screen is wide. Fix missing states and layout breaks, one component at a time, without changing my design and without breaking anything. Summarize by file: what happened before vs. how it works now."
If you still think Claude Code and Codex ๐๐๐ฐ๐ธ ๐ฎ๐ ๐ฝ๐ผ๐น๐ถ๐๐ต, try this:
"Find all the happy-path debt in my app: every screen that only works when data loads fast, nothing fails, lists aren't empty, text is short, and the screen is wide. Fix missing states and layout breaks, one component at a time, without changing my design and without breaking anything. Summarize by file: what happened before vs. how it works now."
Agents build for the demo, where everything goes right. In real use, it doesn't.
1
248
How to cut your cost per PR without losing quality:
Your agent uses the same expensive model to plan a feature and to rename a variable, so even simple work bills at top prices.
Let a strong model plan and a cheaper model carry it out. Zach Lloyd, founder of Warp, says tuning their model mix and factory workflow took cost per PR from $80 to $20, with a lower revert rate and 40% higher PR throughput.
> Step 1: Set your main session to a strong model with /model, like Opus 5.5. That's the model that plans the feature.
> Step 2: Create .claude/agents/executor.md in your repo so a cheaper model makes the edits:
---
name: executor
description: Use proactively for plan steps
tools: Read, Edit, Grep
model: haiku
---
Make the edits for the step you are given. Never change the plan or touch files outside it.
> Step 3: Ask Claude for a plan, then tell it to "use the executor subagent for each step."
Compare your cost per PR before and after, and watch for reverts.
teapawt.org/zachlloydtweets/statusโฆ
At Warp, we are having fun tweaking our model mix and factory workflow to drive our cost per PR down. Itโs gone from $80 โ $20 using a factory approach, with a lower revert rate and 40% higher PR throughput than in the last month.
Folks on the team are suggesting experiments to improve further, like reducing subagents and building model routers. As a startup founder who cares deeply about efficiency and burn, seeing us reduce costs while increasing velocity and quality is incredible.
1
172
How to stop a long-running agent from failing at step 95 of 100:
1. Use a simple retry loop for short-running agents, since each LLM call is isolated, cheap, and stateless.
2. Stop blindly retrying long-running agents, because rerunning hours of work costs API spend, tool calls, and wall clock time, and it can repeat side effects like resending notifications.
3. Plan for failures that aren't AI problems at all: LLM provider rate limits, network timeouts on downstream calls, pods evicted by OOM kills mid-run, and pod rotation during deploys.
4. Plan for context window overflow and human-in-the-loop pauses that can last for days.
5. Add checkpointing so a failed run picks up from its last good step instead of starting over.
6. Make each step idempotent so a retry can't fire the same side effect twice.
7. Keep a durable event log of what the agent has already done.
8. Build resumable workflows, and apply the same backend fundamentals you'd use for any production system.
System design is not important anymore - if you believe this, just wait until your 2-hour agent run fails at step 95 of 100 :) I hope it does not happen, but it actually might.
Agentic apps have two classes: short-running and long-running. When you are building short agents, a simple retry loop is enough. Each LLM call is isolated, cheap, and stateless. Fail, retry, done.
But retrying a long-running agent spanning hours is expensive in API costs, tool calls, and wall clock time, and retries can also have side effects like resending notifications. Some failures that a long-running agent might hit in production are:
- LLM provider rate limiting
- Network timeouts for downstream calls
- Pod eviction due to OOM kills mid-run
- Pod rotation during deployment
- Context window overflow
- Human-in-the-loop pauses for days
Notice how most of these are not AI problems. They are classic software engineering problems that are showing up in a new system.
This is where we need checkpointing, idempotency, durable event logs, and resumable workflows. These are backend engineering fundamentals where we built robust, fault-tolerant systems by applying hundreds of patterns.
Good software engineering practices do not go out of fashion. Because someday, you will ship your "tutorial" to production. That is when system design suddenly matters again.
Fundamentals remain fundamentals, even when the abstraction layer changes.
11
4
319
Grab this repo when you need your Mac awake while your coding agents work, but not forever.
Set a timed session from the menu bar. It can keep agents going with the lid closed while it watches battery and temperature.
Github: github.com/krishhgg/Insomnia
1
1
164
How to find outdated rules in your CLAUDE.md with one Claude Code command:
1. Run /doctor prompt-audit in your repo.
2. Know what it reads: your CLAUDE.md, CLAUDE.local.md, and AGENTS.md files, plus the rules, skills, commands, subagents, and output styles under .claude/ and ~/.claude/.
3. Look for the three problems it flags: instructions written for older models, references to files or commands that don't exist, and files that contradict each other.
4. Pass a path to audit just one file or folder, like /doctor prompt-audit .claude/skills/deploy, or use "." to keep it to the repo you're in.
5. Read the report and proposed edits first, because nothing in your files changes until you ask Claude to apply them.
6. Run it on a stronger model. In one independent test with nine planted problems, Opus 5.5 and Sonnet 5.5 caught all nine in every run, while Haiku 4.5 missed some.
7. Run it again every time you switch models or change your build commands, since those changes turn correct instructions into outdated ones.
1
243
A useful repo if you want to speak your prompts to coding agents instead of typing them.
Hold a key, talk, and your words are pasted into whatever app you're in. It transcribes on your device, on Mac, Windows and Linux.
181
How to run a full UI/UX audit with Fable 5.2:
1. Tell Fable 5.2 to act as a senior product designer with real authority over the codebase, so it can redesign and restructure instead of just patching.
2. Have it derive a taste baseline before it touches anything: who uses the product, for what, and what character it should have. Tell it to treat your existing design docs, tokens, and components as hypotheses to test, not law.
3. Make it walk the core user paths in a real browser at desktop, tablet, and phone widths.
4. Have it fix the correctness floor unconditionally, including overlap, overflow, contrast, 200% text zoom, keyboard nav, visible focus, touch target size, and missing loading, empty, error, and disabled states.
5. Ask it to use its taste on what works but feels cheap, like no focal point, a drifting type scale, broken spacing rhythm, inconsistent radii, and copy that explains the system instead of helping the user act.
6. Give it three levels of authority. It just does small spacing, type, color, and copy fixes. It does section-level restructures as one revertible commit each, with before and after screenshots. It only proposes navigation changes, feature removals, or a new design language.
7. Let it revise your design docs too if the audit shows the design system itself is the problem, and have it cite the page that triggered each change.
8. Ask for a report ranked by impact, with the taste baseline, what it improved and why, the commit evidence, its proposals, and a correctness table of page, viewport, issue, severity, and status.
your claude subscription now has ๐ผ๐ป๐ฒ ๐ณ๐ฟ๐ฒ๐ฒ ๐ณ๐๐น๐น ๐ฟ๐ฒ๐๐ฒ๐, plus the free fable resets. every model limit, back to full in one go.
now's the perfect time to have fable 5.2 run a full ui/ux audit on your product, with the model acting as ๐ฎ ๐๐ฒ๐ป๐ถ๐ผ๐ฟ ๐ฑ๐ฒ๐๐ถ๐ด๐ป๐ฒ๐ฟ ๐๐ถ๐๐ต ๐ฟ๐ฒ๐ฎ๐น ๐ฎ๐๐๐ต๐ผ๐ฟ๐ถ๐๐.
full prompt ๐
act as a senior product designer with real authority over this codebase: you can redesign and restructure, not just patch. your own design judgment is the primary tool.
derive a taste baseline before touching anything:
1. infer product intent in 3 lines: who uses this, for what, what character (tool? editorial? playful? premium?)
2. read the existing design docs, tokens and components as hypotheses to test, not law
3. from 1+2, derive what this product should feel like at its best. judge everything against that, not against trends. gradients, glassmorphism, emoji, decorative motion only if the product's character earns them
walk the core user paths in a real browser at desktop, tablet and phone widths. evaluate on two levels:
A. correctness floor, fix unconditionally: overlap, overflow, clipping, horizontal scroll, contrast, 200% text zoom, keyboard nav, visible focus, modal focus trap and return, labels and roles, touch target size, missing loading/empty/error/disabled states, layout shift
B. design ceiling, use taste to find what works but feels cheap: no focal point, elements fighting for the same weight, copy that explains the system instead of helping the user act, border+shadow+background stacked on one element, drifting type scale, broken spacing rhythm, inconsistent radii, lifeless interactions
three levels of authority:
L1 just do: spacing, type, color token fixes, copy trims, micro-interactions, deleting redundancy
L2 do but isolate: section-level restructures, component merges, one revertible commit each with before/after screenshots
L3 propose only: navigation model changes, removing features, changing the design language itself
if the audit shows the documented design system is itself the problem, revise the docs too, citing the specific page that triggered each change.
deliver a report ranked by impact: the taste baseline you derived, what you improved and why, L2 evidence, L3 proposals, and a correctness table (page, viewport, issue, severity, status).
& thank me later
2
357















