# 100 levels of working with agents 100 things people who are good with coding agents actually do, from pasting screenshots to adversarial reviewers. Paste one prompt: your agent checks which you do on your own machine and hands you a card. - Page: https://www.prehistoriq.com/levels/ - Check your own level: paste `Read prehistoriq.com/levels/check.md and do what it says.` into your coding agent - Data: https://www.prehistoriq.com/levels/levels.json · Season 1, checked 2026-10-21 Your level is how many of the 100 you do. The list is always 100: when a new level earns its place, one goes extinct. Level 0: Copy code out of a chat tab. Everyone starts here. It doesn't count, and nobody's judging. ## Levels 1–10: Stone Age · First contact It edits, it runs things, it answers. You stop copy-pasting. ### 1. Install a coding agent One that reads your files and runs commands, not a chat tab. - How it's checked: Any agent config folder exists: ~/.claude, ~/.codex, ~/.cursor, ~/.config/opencode, ~/.gemini, ~/.factory, ~/.pi. You're reading this from inside one, so: yes. - Who does it: Mitchell Hashimoto: “To find value, you must use an agent.” (https://mitchellh.com/writing/my-ai-adoption-journey) ### 2. Let it edit the files Stop pasting code between a chat window and your editor. It writes to disk; you read the diff. - How it's checked: Its file-edit tool shows up in session logs (Edit, Write, apply_patch). ### 3. Ask before you change "How does auth work here?" Ask it what you'd ask a senior engineer on day one. - How it's checked: 5+ prompts that are questions about the code (how, why, where, what does). - Who does it: Anthropic, Claude Code best practices: Ask it the questions you'd ask a senior engineer about the codebase. (https://code.claude.com/docs/en/best-practices) ### 4. Paste the whole error The full stack trace or log, not your summary of it. - How it's checked: 5+ prompts with pasted blocks or raw logs ([Pasted text …], stack traces). ### 5. Send it a screenshot A picture of the broken layout beats a paragraph about it. "Fix this" + image is a whole prompt. - How it's checked: 3+ prompts with an image attached ([Image #…]). - Who does it: Peter Steinberger: Attaches a screenshot to a large share of his prompts; "fix padding" plus an image is enough. (https://steipete.me/posts/just-talk-to-it) ### 6. Interrupt it Hit Esc the moment it heads the wrong way. Waiting for it to finish costs more. - How it's checked: 5+ interrupted turns in session logs ("[Request interrupted by user]"). - Who does it: Peter Steinberger: “Don't be afraid of stopping models mid-way.” (https://steipete.me/posts/just-talk-to-it) ### 7. Point at the files @-mention the file or paste the path, so it doesn't go hunting. - How it's checked: 10+ prompts with @file mentions or explicit file paths. ### 8. Commit small and often A clean git tree is the best undo button you have. Commit before you let it loose. - How it's checked: It ran git commit 5+ times in your sessions, or you use checkpoints. ### 9. Start a fresh session When it gets confused, /clear and ask again. Arguing with a polluted context never wins. - How it's checked: /clear, /new or a fresh session after a failed attempt, 3+ times. - Who does it: Anthropic, Claude Code best practices: After two failed corrections, /clear and write a better first prompt. (https://code.claude.com/docs/en/best-practices) ### 10. Use it for something that isn't code Rename 400 photos, sort the receipts, untangle a spreadsheet. It's a computer that does chores. - How it's checked: 1+ session in a folder that isn't a git repo (Downloads, Documents, home). ## Levels 11–20: Fire · Context It knows what you know, because you told it. ### 11. Write a rules file AGENTS.md or CLAUDE.md: how to build, test and run this repo, and what never to do. - How it's checked: An AGENTS.md, CLAUDE.md, GEMINI.md or .cursor/rules in any repo you work in. - Who does it: Lee Robinson: “I started with zero custom rules. I only added a rules file later after the model got a few things wrong” (https://leerob.com/pixo) ### 12. Write down how you like to work One global file for every project: your taste, your tools, your pet peeves. - How it's checked: ~/.claude/CLAUDE.md or ~/.codex/AGENTS.md exists and isn't empty. ### 13. Talk instead of type Dictate. You'll give it three times the context you'd ever bother typing. - How it's checked: A dictation app is installed (Wispr Flow, superwhisper, MacWhisper, VoiceInk, OpenWhispr, Aqua, Spokenly, Handy), or prompts read like speech (um, uh, long run-ons). - Who does it: Dax Raad: “the thing that killed my editor for me was fast local voice models more than OpenCode” (https://x.com/thdxr/status/2078527322563449007) ### 14. Hand it the docs Paste the URL of the API reference, the issue, the RFC. Let it read instead of guess. - How it's checked: 10+ prompts with URLs, or web fetches in session logs. ### 15. Point at an example "Do it like UserCard.tsx." One good example beats a paragraph of conventions. - How it's checked: Prompts that name an existing file or project to copy ("like X", "same as Y", "follow Z"). ### 16. Say what done looks like What's wrong, where, and how you'll know it's fixed. "Fix login" isn't a task. - How it's checked: Prompts that state a finish line ("done when", "until", "should show"). - Who does it: OpenAI, Codex best practices: Goal, Context, Constraints, Done when: their prompt template. (https://learn.chatgpt.com/guides/best-practices) ### 17. Ask for a plan first Plan mode, or "don't write code yet". A plan is ten times shorter than the diff it saves you. - How it's checked: Plan mode used, or 5+ prompts asking for a plan before code. - Who does it: Lee Robinson: “Plan mode in Cursor is one of the most underrated features.” (https://leerob.com/pixo) - Who disagrees: Peter Steinberger: Skips plan mode; asks for options and talks it through instead. (https://steipete.me/posts/just-talk-to-it) ### 18. Ask for three versions Of the UI, the name, the approach. Pick one, or steal from all three. Variants are nearly free now. - How it's checked: 5+ prompts asking for variants, options or versions. ### 19. Let it interview you "Ask me questions until you understand what I want." Then it builds the right thing. - How it's checked: Its ask-the-user tool used, a grill-me style skill installed, or prompts like "ask me" / "interview me". - Who does it: Matt Pocock: grill-me (https://github.com/mattpocock/skills/blob/7a030a8b0dcaa601bb0dda9aaf174affb3d61145/README.md#L103) ### 20. Add a rule after every mistake It did something dumb? It goes in the rules file, so it never does it twice. - How it's checked: A rules file changed 3+ times, agent memory notes saved, or prompts like "add this to AGENTS.md". - Who does it: Boris Cherny: “Anytime we see Claude do something incorrectly we add it to the CLAUDE.md” (https://x.com/bcherny/status/2007179840848597422) ## Levels 21–30: Wheel · Checks It proves the thing works before it says done. ### 21. Give it a way to check its work Tests, a build, a screenshot to compare. Then it fixes its own mistakes before you see them. - How it's checked: 10+ prompts that ask it to verify, test, or make sure it works. - Who does it: Boris Cherny: “give Claude a way to verify its work… 2-3x the quality” (https://x.com/bcherny/status/2044847858634064115) ### 22. Failing test first Reproduce the bug in a test, watch it fail, then fix. Proof before the fix. - How it's checked: Prompts about failing tests, TDD or red/green, or a TDD skill installed. - Who does it: Steph Ango: “Verify the same fixture fails before the fix and passes afterward; do not change expectations to match buggy behavior.” (https://github.com/kepano/defuddle/blob/49d14ecc4089696565c2479b2f57db72be199adc/AGENTS.md#L17) ### 23. Let it drive a browser Playwright, agent-browser, Chrome. It clicks through what it built instead of guessing. - How it's checked: A browser MCP or skill installed, or 10+ browser-automation calls. ### 24. Ask for proof, not "done!" The command it ran and the real output. A screenshot. A video of it working. - How it's checked: 5+ prompts asking for proof, screenshots or recordings of the result. - Who does it: Thorsten Ball: “give me irrefutable proof that this works. Dazzle me. Give me a video as proof, or something.” (https://registerspill.thorstenball.com/p/joy-and-curiosity-99) ### 25. Undo the conversation, not just the code Rewind or a checkpoint takes you back to before it went wrong. Try risky things. - How it's checked: /rewind or a checkpoint restore used by you at least once (typed, not just mentioned). ### 26. Teach it your CLIs gh, aws, stripe, supabase. "Run foo --help" and it learns the rest. - How it's checked: It ran 3+ different third-party CLIs (gh, aws, gcloud, vercel, netlify, stripe, supabase, wrangler, fly…). ### 27. Pre-approve the safe stuff An allowlist for tests, lint and git status. Approval fatigue is how you end up clicking yes to everything. - How it's checked: Permission allow rules, Codex rules, or auto mode configured. ### 28. Ask side questions without derailing /btw, or a second session. The answer never pollutes the main context. - How it's checked: /btw used, or a parallel throwaway session for a quick question. ### 29. Type the next message while it works Queue follow-ups instead of waiting for it to finish. - How it's checked: Queued prompts in your agent's stats or logs, 20+. ### 30. Get pinged when it's done A hook, a sound, a push to your phone. Then go do something else. - How it's checked: A hook or script that notifies you when it's done (sound, push, notification), or a rule telling it to. - Who does it: Chris McCord: “when it's done work it runs a `notify msg` bash function I have” (https://x.com/chris_mccord/status/2016576023400181773) - Who disagrees: Mitchell Hashimoto: Turns agent notifications off, so he decides when to switch context. (https://mitchellh.com/writing/my-ai-adoption-journey) ## Levels 31–40: Bronze Age · Tools It reaches past the repo: CLIs, servers, your browser, your inbox. ### 31. Install a skill Someone else's packaged know-how, loaded only when it's needed. - How it's checked: 1+ skill folder in ~/.claude/skills, ~/.codex/skills, ~/.agents/skills or a repo's .claude/skills. ### 32. Connect a tool server MCP for the things without a good CLI: Figma, Linear, your database. - How it's checked: 1+ MCP server configured. - Who disagrees: Peter Steinberger: Prefers CLIs: the GitHub MCP alone ate 23k tokens of context. (https://steipete.me/posts/just-talk-to-it) ### 33. Let it open the PR Branch, commit, push, write the description. You review. - How it's checked: It ran gh pr create (or similar) 3+ times. ### 34. Let it ship Vercel, Netlify, Fly, a TestFlight build. You decide who presses the last button. - How it's checked: It ran a deploy CLI 3+ times. ### 35. Save a prompt as a command The thing you type every week, saved as /ship or /review. - How it's checked: 1+ custom command in .claude/commands, or a user-invoked skill you wrote. ### 36. Turn the effort dial Low for chores, high for the gnarly bug. Pick the model per job. - How it's checked: /effort or /model switched 3+ times, or reasoning effort set in config. ### 37. Know what you burn ccusage, CodexBar, /usage. Tokens are a budget; spend them on purpose. - How it's checked: A usage tracker installed, or /usage, /stats, /cost used 3+ times. ### 38. Watch the context fill up A status line with context used, model and cost. No more surprise compactions. - How it's checked: A statusLine configured, or /context used 3+ times. ### 39. Two agents at once One on the bug, one on the feature. Different folders, different problems. - How it's checked: Two sessions active in the same 10 minutes, on 5+ occasions (from prompt timestamps). ### 40. Connect it to your messages Gmail, Slack, Teams, calendar. "What did I promise people this week?" - How it's checked: A mail, chat or calendar connector or MCP configured (names only). ## Levels 41–50: Iron Age · Your setup Rules, skills, hooks and memory that you wrote yourself. ### 41. Write your own skill The thing you keep explaining, written down once as a SKILL.md. - How it's checked: A skill you wrote yourself: no upstream source or lockfile entry, or it lives in your own repo. ### 42. Use subagents Send helpers to read and search, so your main context stays clean. - How it's checked: 10+ subagent runs in session logs. - Who does it: Alexander Embiricos: “Then, when I need anything new, I say "using a subagent in parallel, do X"” (https://x.com/embirico/status/2046033614106931414) ### 43. A hook that tidies every edit Formatter or linter after each write. Deterministic beats "please run prettier" in a rules file. - How it's checked: A post-edit hook that runs a formatter or linter. - Who does it: Boris Cherny: “We use a PostToolUse hook to format Claude's code.” (https://x.com/bcherny/status/2007179852047335529) ### 44. A hook that says no Block rm -rf, force pushes, pip when you use uv. Rules are advice; hooks are law. - How it's checked: A pre-tool hook that blocks or guards commands. - Who does it: Armin Ronacher: “Replaces the bash tool with a `uv`-aware version that injects Python command shims and blocks common non-`uv` workflows.” (https://github.com/mitsuhiko/agent-stuff/blob/0865c849befd2021490679f96a8dee58c84ac857/README.md#L63) ### 45. Let it keep notes Memory, notes.md, PROGRESS.md. It shouldn't relearn your project every morning. - How it's checked: 5+ memory notes, Codex memories on, or notes/progress files in your repos. - Who does it: Simon Willison: “Create a notes.md file in that folder and append notes to it as you work, tracking what you tried and anything you learned along the way.” (https://github.com/simonw/research/blob/ba0cb97b42da153c13c465fb617f5c45f1f03b04/AGENTS.md#L3) ### 46. One rules file for every agent Symlink CLAUDE.md to AGENTS.md. Same rules, whichever agent you open. - How it's checked: A CLAUDE.md that is a symlink to, or imports, an AGENTS.md (or the other way round). - Who does it: Peter Steinberger: “~/.codex/AGENTS.md -> ~/Projects/agent-scripts/AGENTS.MD ~/.claude/CLAUDE.md -> ~/Projects/agent-scripts/AGENTS.MD” (https://github.com/steipete/agent-scripts/blob/79150cfac4a6ba4df13cf30176bf2db85170ae6c/README.md#L65-L67) ### 47. Keep the rules file small A map to your docs, not an encyclopedia. Would removing this line cause a mistake? No? Delete it. - How it's checked: The rules file of the repo you work in most is under 150 lines and points to other docs for the rest. - Who does it: Kent C. Dodds: “This file is included in full context for every agent conversation. Keep it tiny and stable.” (https://github.com/kentcdodds/kentcdodds.com/blob/b61c7ff4200c4aa0e35326d1a7baeb8a51d86673/AGENTS.md#L3-L4) ### 48. Spec first Two paragraphs in, a SPEC.md out, and a fresh session builds from the spec. - How it's checked: A SPEC.md, PLAN.md or plans/ folder in a repo you work in. - Who does it: Alex Kladov: “I was providing structure, Claude was painting by the numbers.” (https://matklad.github.io/2026/01/20/vibecoding-2.html) ### 49. Write a handoff Before the context fills, it writes down where it is. The next session picks up in one read. - How it's checked: A PROGRESS.md, HANDOFF.md or todo.md, or a handoff skill. - Who does it: Thomas Ptacek: “I should be able to tell a future agent "read PLAN.md and PROGRESS.md" and trust it's up to speed with this codebase.” (https://github.com/tqbf/swiftui-app/blob/3ca1cc4b8ffdbd3b628cc847574c9951b036fe12/CLAUDE.md#L5-L6) ### 50. One worktree per agent Parallel agents in one checkout trip over each other. Worktrees don't. - How it's checked: git worktree list shows 2+ in any repo. - Who does it: Boris Cherny: “Worktrees are essential… dozens of Claudes” (https://x.com/bcherny/status/2038454353787519164) - Who disagrees: Peter Steinberger (2025, he uses them now): “Why not worktrees? I run one dev server, as I evolve my project I click through it and test multiple changes at once.” (https://steipete.me/posts/just-talk-to-it) ## Levels 51–60: Steam · Second opinions Another agent reads it before you do. ### 51. Have an agent review the diff /code-review, Codex /review, a fresh session. Before you read it, not instead. - How it's checked: A review command, review skill or review subagent used 3+ times. - Who does it: Filippo Valsorda: “making a human review a change before an LLM is disrespectful.” (https://bsky.app/profile/filippo.abyssdomain.expert/post/3moq6sdq6xk2l) ### 52. Cross-model review Claude writes, GPT reviews, or the other way round. Different models miss different things. - How it's checked: One vendor's agent asked to review another's work (codex exec / claude -p review, or prompts that hand a diff across). - Who does it: DHH: “Review this with claude” (https://x.com/dhh/status/2107882896195199116) ### 53. Fresh-eyes review Same model, empty context: "read this as if you didn't write it". - How it's checked: Reviews run in a fresh context (a new session or a subagent), not in the session that wrote the code. - Who does it: Harper Reed: “carefully read over all of the new code you just wrote … with "fresh eyes," looking super carefully for any obvious bugs … Use a subagent.” (https://github.com/harperreed/dotfiles/blob/6db272a76a71b098272c7c8a26ae084d197a43ac/.claude/commands/careful-review.md#L1) ### 54. Make it quiz you on the diff It explains the change, then asks you five questions. Merge only if you pass. - How it's checked: An explain-diff or quiz skill, or prompts asking it to explain or quiz you on a change. - Who does it: Thariq Shihipar: “I only merge after I pass the quiz perfectly.” (https://x.com/trq212/status/2073100352921215386) ### 55. Ask for a diagram Architecture, data flow, the state machine. Pictures review faster than prose. - How it's checked: 3+ prompts asking for a diagram, chart or visual explanation. - Who does it: Dex Horthy: “/show-me prompts the agent to use concise visuals to explain what's happening, instead of walls of prose.” (https://www.humanlayer.dev/blog/show-me-skill) ### 56. Let it babysit the PR It watches CI and the review bots, fixes, pushes, and repeats until it's green. - How it's checked: It checked CI or PR status (gh pr checks, gh run) 5+ times. ### 57. Two agents, two jobs Claude writes the code, Codex draws the images. Or one builds and the other reviews. - How it's checked: One agent launched another (codex exec, claude -p, gemini -p) 3+ times. ### 58. Run it headless claude -p, codex exec. Pipe logs in, get JSON out, put it in a script. - How it's checked: 3+ headless runs in shell history or logs. ### 59. Have it make a video A 30-second motion explainer of what you shipped, for the colleague who won't read the PR. - How it's checked: Remotion, Hyperframes or ffmpeg renders from a session, or a video skill installed. ### 60. Skip the prompts, inside a box Full autonomy in a container or a fresh VM with no prod keys. Not in your home folder. - How it's checked: Permission prompts off AND a sandbox (devcontainer, Docker, VM, /sandbox). Prompts off without a box doesn't count. - Who does it: David Crawshaw: “You have to turn off the sandbox, which means you have to provide your own sandbox. I have tried just about everything and I highly recommend: use a fresh VM.” (https://crawshaw.io/blog/eight-more-months-of-agents) ## Levels 61–70: Electric · Off the desk It works while you sleep, sit on the couch, or close the laptop. ### 61. Schedule a prompt For later, or every morning: /loop, /schedule, cron, launchd. - How it's checked: A scheduled or recurring agent run (loop/cron tools, a launchd or cron entry that calls an agent). ### 62. Let it work while you sleep Kick off a few tasks before bed, review them with coffee. - How it's checked: Agent activity between 1am and 6am while you typed nothing. - Who does it: Calvin French-Owen: “This is over-hyped on Twitter, but I do really try and kick off 3-4 tasks in Codex before going to bed, so they'll be ready for review in the morning.” (https://calv.info/agents-feb-2026) ### 63. Drive it from your phone Remote Control, the Codex app, ssh + mosh + tmux. Unblock it from the couch. - How it's checked: Remote control set up, a mobile session, or mosh/Tailscale used to reach your agent box. - Who does it: Harper Reed: Claude Code is better on your phone (https://harper.blog/2026/01/05/claude-code-is-better-on-your-phone/) ### 64. Hand chores to a cloud agent Codex cloud, Claude on the web, Cursor background agents, Devin. The boring stuff, no laptop needed. - How it's checked: Cloud or remote agent tasks in your history. - Who does it: Adam Wathan: “I feed this thing every little random tedious OSS task that I'd otherwise neglect on my TODO list” (https://x.com/adamwathan/status/2017616165455949851) ### 65. A finish line it can't fake "Don't stop until the test passes and the screenshot matches." Then let it run. - How it's checked: 3+ prompts like "don't stop until" / "keep going until", or a goal set with /goal. ### 66. Run it in a loop Same prompt, fresh context, again and again until the spec is done. - How it's checked: A shell loop or script that re-runs an agent, or /loop until a condition. - Who does it: Geoffrey Huntley: “while :; do cat PROMPT.md | claude-code ; done” (https://ghuntley.com/ralph/) ### 67. Put an agent in CI @claude on issues, Codex on every PR. It's a teammate with a GitHub login now. - How it's checked: claude-code-action, codex action or similar in a .github/workflows file. ### 68. Run it four times, keep the best Same task, parallel attempts, thirty seconds each to judge. Throwing work away is fine now. - How it's checked: The same task run in 2+ parallel sessions or worktrees. - Who does it: Alexander Embiricos: “So I might kind of throw throw codex at it but I might just do it four times in parallel instead.” (https://www.youtube.com/watch?v=xeZDHGjG5zM&t=1375) ### 69. Make two agents argue One proposes, one attacks, until they agree or you see exactly why they can't. Nice. - How it's checked: Multi-round back-and-forth between two agents (rounds of review and fix, or relayed replies). - Who does it: Dex Horthy: “let sol and fable implement / review / argue over night” (https://x.com/dexhorthy/status/2092244659057909765) ### 70. An always-on machine A box that keeps agents running when your laptop sleeps. A Mac mini, a VPS, a shed. - How it's checked: Agents running on another machine you reach over ssh. - Who does it: DHH: “all my agent and compile work now happens on an AI shed.” (https://x.com/dhh/status/2107463006552793231) ## Levels 71–80: Silicon · Orchestra Many agents, one plan, and you still know who's doing what. ### 71. Five at once Five agents, five problems, and you still know what each one is doing. - How it's checked: 5+ sessions or background agents running at the same time. ### 72. A coordinator agent One agent hands out the work, checks on the others, and reports to you. - How it's checked: One agent messaging or steering others (agent teams, SendMessage, an orchestrator session). - Who does it: Andreas Kling: “Tell coordinator to parallelize the work over 4 separate computers via SSH” (https://x.com/awesomekling/status/2102089363631059199) ### 73. Strong model plans, fast model builds Pay for thinking where it matters. Different models for design, code and review. - How it's checked: Per-agent model settings, or 3+ prompts assigning models to roles. - Who does it: Steve Yegge: “Every implementation bead goes through this lifecycle: Fable design, Opus implementation, Fable review.” (https://yegge.ai/essays/the-shape-of-things-to-come/) ### 74. Give agents job titles A security reviewer, a tester, a docs writer, each with its own tools and model. - How it's checked: 2+ custom agent definitions (.claude/agents/*.md or equivalent). ### 75. Script the whole pipeline A rerunnable multi-agent workflow, not a heroic one-off session. - How it's checked: A workflow tool used, or a script that chains several agent runs. ### 76. Adversarial reviewers Two refuters per builder, and points for every real bug they find. - How it's checked: Reviews run by 2+ competing or adversarial agents. - Who does it: Jesse Vincent: “Assign two subagents to review your work adversarially. Tell them that whoever finds the largest number of legitimate significant issues gets 5 points.” (https://github.com/obra/dotfiles/blob/6ebb5ba848526b2bfd5b5ab3cb30741ff32c4447/.claude/commands/par.md#L1) ### 77. Three bug hunters, one report Run several reviewers, then have one agent rule out the false positives. - How it's checked: Findings from 2+ review tools merged into one triaged list. - Who does it: Nolan Lawson: “Run a Claude sub-agent, Codex, and Cursor Bugbot to find bugs in this PR ranked by critical/high/medium/low.” (https://nolanlawson.com/2026/05/25/using-ai-to-write-better-code-more-slowly/) ### 78. Every task makes the next one easier After each job it writes the lesson into a skill or rule. Next time is cheaper. - How it's checked: Rules, skills or memory edited by agent sessions 5+ times. - Who does it: Kieran Klaassen: “Compound is still the most important step, because the whole point is that every feature should make the next feature easier.” (https://every.to/p/compound-engineering-gets-an-upgrade) ### 79. Turn rules into checks Written "never do X" twice? Make it a lint rule or a type, so it can't happen. - How it's checked: Custom lint rules, ast-grep rules or check scripts in a repo. - Who does it: José Valim: “encode as many rules as possible into deterministic checks (first) and adversarial reviews (second), instead of relying exclusively on a bunch of AGENTS.md files to do the job.” (https://x.com/josevalim/status/2075221914549686774) ### 80. Build a tool for your agent A small CLI made for it to call: --json output, good --help, no prompts. - How it's checked: Your own scripts or CLIs that agents call regularly. ## Levels 81–90: Orbit · Harness You build the system the agents work in. ### 81. Test your skills Does it trigger when it should? Does it actually help? An eval tells you. - How it's checked: Eval files for a skill or plugin, or a plugin eval run. ### 82. Publish your skills A public repo of your skills, so your team (or strangers) can steal them. - How it's checked: Your skills live in a git repo with a remote. - Who does it: Peter Steinberger: agent-scripts (https://github.com/steipete/agent-scripts) ### 83. An agent in your team's chat Tag it in Slack, it opens the PR. Your teammates use it without installing anything. - How it's checked: A chat-integrated agent set up (Slack app, Claude Tag, Devin, a bot). - Who does it: Cat Wu and Thariq Shihipar, via Simon Willison: Claude Tag lands 65% of the Claude Code team's product PRs. (https://simonwillison.net/2026/Jul/21/cat-and-thariq/) ### 84. Build on the Agent SDK Your own agent product, with the same loop under the hood. - How it's checked: An agent SDK in a project's dependencies (Claude Agent SDK, Codex SDK, OpenAI Agents). ### 85. A wiki about you, for your agents How you think, what you've decided, what you can't stand. Loaded only when it's relevant. - How it's checked: A personal context repo or folder that your agents load when relevant, through a rule or a skill. ### 86. Your agents talk to each other No more "Codex says: [paste]". They pass findings directly while you do something else. - How it's checked: Agents sending each other messages or reading each other's output, without you relaying it. ### 87. Sensors next to the agent Lint, types, coverage, mutation score, always running, where it can read them. - How it's checked: Watch-mode checks or a dashboard the agent reads while it works. - Who does it: Birgitta Böckeler: Sensors for coding agents (https://martinfowler.com/articles/sensors-for-coding-agents.html) ### 88. A QA script it runs before release Every release, the same long checklist, run by the agent, start to finish. - How it's checked: A QA or release checklist file, or a release skill. - Who does it: Salvatore Sanfilippo: QA_BEFORE_RELEASES.md (https://antirez.com/news/168) ### 89. Hardware in the loop It flashes the board, reads the serial port, watches the LED. Or the Stream Deck. - How it's checked: Device tools driven by an agent (serial, adb, simctl, esptool, arduino-cli…). - Who does it: Linus Torvalds: “I started the day with "I want to connect two pedals together and have them validate each other", and ended the day with three pedals, a signal generator, and a USB MIDI dongle all connected together as a validation bench with automated checks.” (https://github.com/torvalds/GuitarPedal/commit/d1b95dc5e71e0a2d42912001c5fa4469bbc88766) ### 90. Your git log obeys your rules file If the rules say "no AI trailers", the last 100 commits have none. If they say "tests pass", they did. - How it's checked: Compare what your global rules demand with your last 100 commits (trailers, messages). They agree. - Who does it: Mike McQuaid: “reject "has AI attribution; remove it and disclose AI use in the PR description."” (https://github.com/MikeMcQuaid/dotfiles/blob/a340d0ef58084694861946d587a0832a80782898/git-hooks/pre-push#L51) ## Levels 91–100: Stone, again · Full circle Less config, more judgement. You read the 5% that matters. ### 91. Delete more config than you add Prune the rules, skills and servers that don't earn their tokens anymore. - How it's checked: A rules file that got shorter at least once in its git history. - Who does it: Dax Raad: “when i see people with custom workflows and setups they're all addressing problems that don't exist anymore” (https://x.com/thdxr/status/2107260844400931156) ### 92. Start clean with each new model Old prompts fight new models. Reset, then add back only what's missing. - How it's checked: Your setup was trimmed or rewritten soon after a major model release. - Who does it: Cat Wu: “I generally reset my configs for each new model” (https://x.com/_catwu/status/2102556970473709717) ### 93. Fewer agents, on purpose You've run ten at once, and you know exactly why you run two now. - How it's checked: You used to run more parallel agents than you do now, by choice. - Who does it: Thariq Shihipar: “now like it's mostly one active Claude Code session and then a bunch of Claude Tag sessions.” (https://www.youtube.com/watch?v=aVO6E181cNU&t=1200s) ### 94. Write the skeleton yourself Types, signatures, the hard 20%. It fills in the bodies. - How it's checked: You write structure and interfaces by hand before handing over. - Who does it: Alex Kladov: “I was providing structure, Claude was painting by the numbers.” (https://matklad.github.io/2026/01/20/vibecoding-2.html) ### 95. Review by risk, not by habit A copy fix gets a glance. Auth, money and migrations get every line. - How it's checked: Your rules or review setup name the paths that always get a human read. - Who does it: Boris Cherny: “I read diffs when it's touching critical code, or code that's important to get right. I don't write the code.” (https://x.com/bcherny/status/2098210496109572123) ### 96. Cap what it writes per day No more code than you can actually review. The bottleneck is you, so be the bottleneck on purpose. - How it's checked: You hold back generation to what you can review. - Who does it: Mario Zechner: “Set yourself limits on how much code you let the clanker generate per day, in line with your ability to actually review the code.” (https://mariozechner.at/posts/2026-03-25-thoughts-on-slowing-the-fuck-down/) ### 97. Teach it to someone Pair with a colleague, run a workshop, write the post. You don't know a trick until you've taught it. - How it's checked: Only you know. ### 98. It knows when to stop and ask Stuck agents ask instead of trying a ninth fix. You set that up. - How it's checked: A rule or hook that makes it stop and report when stuck or out of scope. ### 99. Read all 3,000 lines, once At least once, on something that mattered, you read every line it wrote. Then you know what to skip. - How it's checked: Only you know. - Who does it: Mitchell Hashimoto: “First, I reviewed the entire diff (yes, all ~3,000 lines of it). Then I reviewed it again. I re-read the entire diff by hand three or four times.” (https://mitchellh.com/writing/simdutf-no-libcxx) ### 100. Read the right 5% Know which part of the diff can hurt you, and read that part properly. Every time. - How it's checked: Only you know. ## Fossil record - "You are a senior engineer with 20 years of experience." (extinct, 2023): Role-play preambles. Today's models don't need the pep talk; they need the context. - "I'll tip you $200 for a perfect answer." (extinct, 2023): It sort of worked once. Nobody ever paid. - "Take a deep breath and work on this step by step." (extinct, 2023): Reasoning models think before answering anyway. - "Write the full code. No placeholders." (extinct, 2023–24): From the era of lazy models and "// rest of the code stays the same". - .cursorrules (extinct, 2024): One file in the repo root. Replaced by rule folders and AGENTS.md, which every agent reads. - Codex custom prompts (extinct, 2025): Deprecated in favour of skills. - "ultrathink" (endangered, 2025–): Thinking keywords. Effort is a setting now, and reports disagree on whether the word still does anything. - "How many r's are in strawberry?" (extinct, 2024): Retired as a benchmark. Still a good party trick.