Banning AI words one by one never ends: 6 rules from 1946 for your global rules file, plus prompts for docs, PRs and reports.
Writing rules, from Orwell, 1946. These govern prose: docs, PR text, messages. Never touch code or technical terms; swap in everyday words only where precision survives.
1. Never use a metaphor, simile or other figure of speech which you are used to seeing in print.2. Never use a long word where a short one will do.3. If it is possible to cut a word out, always cut it out.
Where a long task stands while you wait: an HTML dashboard of progress, blockers and open questions, built by a side subagent.
Set up a subagent that only builds progress dashboards:
1. Create dashboard-builder in ~/.claude/agents: model: opus, effort: medium, memory: user. Preload the design skills I have installed (for example impeccable). It can only read and write the .dashboard/ folder and its own memory.2. The first time, it asks me what style I like: dark or light, dense or airy, and one accent color. It saves my answer to memory…
Most saved in this topic, as of 22 Oct 2026Most saved in this topic, as of 29 Oct 2026Most saved in this topic, as of 29 Oct 2026Most saved in this topic, as of 29 Aug 2026Most saved in this topic, as of 24 Oct 2026Most saved in this topic, as of 24 Oct 2026Most saved in this topic, as of 16 Aug 2026
Fable quota idle while Opus 5.5 does the work: Fable called in before a plan, when an error repeats and before ‘done’.
Prompt
1Set up my Claude Code to match this tree:21. For the three subagents, reuse fitting ones from ~/.claude/agents and .claude/agents, and propose new ones only for missing roles. Set each to model: opus and effort: medium. Leave any subagent that already sets a different model alone, and list it.32. Set the main session's default effort to high with effortLevel in ~/.claude/settings.json.43. Check for environment variables that turn the advisor off (CLAUDE_CODE_DISABLE_ADVISOR_TOOL, or DISABLE_TELEMETRY and anything else that stops flag fetching) and for CLAUDE_CODE_EFFORT_LEVEL, which overrides every subagent's effort. Report them; don't change them.54. Add one rule to ~/.claude/CLAUDE.md: ask the advisor before committing to a big plan, when the same error shows up a second time, and before calling a long task done.6Show me the changes first. Don't edit any files yet.
Bounded delegated work at max reasoning: a luna_worker custom agent, checked against your Codex version, diff shown first.
Prompt
1create a custom agent named luna_worker at ~/.codex/agents/luna-worker.toml.2use these settings:3model = "gpt-5.6-luna"4model_reasoning_effort = "max"5give it a description and instructions for bounded delegated work. preserve the rest of my config. validate it against my installed Codex version, show me the diff, then use luna_worker for subagent tasks.
Where a long task stands while you wait: an HTML dashboard of progress, blockers and open questions, built by a side subagent.
Prompt
1Set up a subagent that only builds progress dashboards:21. Create dashboard-builder in ~/.claude/agents: model: opus, effort: medium, memory: user. Preload the design skills I have installed (for example impeccable). It can only read and write the .dashboard/ folder and its own memory.32. The first time, it asks me what style I like: dark or light, dense or airy, and one accent color. It saves my answer to memory and follows it every time. Pick the panels for the current task; don't use a template.43. The dashboard is one HTML file showing tasks and their status, questions waiting for me with the default action, the latest deliverables, and anything stuck. Use the real clock for every time. It opens with a double-click and refreshes itself every 10 seconds.54. Add a rule to ~/.claude/CLAUDE.md: for any task with more than 5 steps or that should take longer than 30 minutes, have dashboard-builder set up the dashboard before starting; update it after every step; when you need a decision from me, add it to the questions list and keep going with the default.6Show me the contents of the files you'll create or change, then explain the whole flow in words a 10-year-old could follow. Don't write anything until I confirm.
A vibe-coded site about to go live: design, mobile, every state, real-user flows and launch details, fixed in one pass.
Prompt
1Before you ship a vibe-coded site, have Opus 5.5 send out 5 subagents to check these 20 things, then fix them all in one pass:2Consistent design3- Write a DESIGN.md first, and make every color, font size, spacing and corner radius follow it4- Keep the palette restrained and the type hierarchy clear5- Make every button, card and input of the same kind look alike6- Too flat? Add emphasis. Too crowded? Cut7- Keep text readable on both dark and light backgrounds
Banning AI words one by one never ends: 6 rules from 1946 for your global rules file, plus prompts for docs, PRs and reports.
Global CLAUDE.md / AGENTS.md rules
1Writing rules, from Orwell, 1946. These govern prose: docs, PR text, messages. Never touch code or technical terms; swap in everyday words only where precision survives.21. Never use a metaphor, simile or other figure of speech which you are used to seeing in print.32. Never use a long word where a short one will do.43. If it is possible to cut a word out, always cut it out.54. Never use the passive where you can use the active.65. Never use a foreign phrase, a scientific word or a jargon word if you can think of an everyday English equivalent.76. Break any of these rules sooner than say anything outright barbarous.
A pile of sessions and nothing done: today’s merges told apart from busywork, then one sentence on tomorrow’s first task.
Prompt · Claude Code, Codex · Review ·
Don't summarize yet, and don't praise me. Ask me one question at a time, and wait for my answer before the next one.
1. First, look at today's commits, branches and uncommitted changes yourself. Pick out what actually got merged and what was just back-and-forth edits. Then ask me: Going only by these records, not by what I say, what would an outsider thin…
Pages that scream AI at first glance: every slop sign listed, a five-line direction for your OK, then fixes pass by pass.
Prompt · Claude Code · Design ·
Take the AI look out of this page. Don't touch features, content or section order. Only change the craft.
1. Audit: list every sign of AI slop and where it is. Start with purple gradients, glows, gradient text, emoji labels, "New ✨" badges, all-caps labels with wide letter spacing, animations that never stop, spacing that's bigger here and smal…
You can’t find your AI-built site by Googling its name: crawling, every page and the whole site, fixed in one pass.
Prompt · Claude Code · Launch + Review ·
Have Opus 5.5 send out 3 subagents to check these 12 things, then fix them all in one pass:
Crawling
- Curl every main page with Googlebot's user agent. The title and body text should be there without JS
Astra flags bugs that almost never happen and Opus fixes each one: bugs scored, low scores filed as issues, code left alone.
Rules · Claude Code, Codex · Review ·
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.
Requests you repeat, commands you stop or undo, how you check changes: ranked by how much each would help, your words quoted.
Prompt · Claude Code · Codebase ·
Then use it to go through my last 20 sessions (this project's .jsonl logs under ~/.claude/projects/) and see how I use you day to day. Find three kinds of things worth turning into a mod:
1. Things I keep asking you, or asking you to do. Each mod in the guide answers one question, like how full the context is. Pick out the questions and requests I repeat acro…
Broken features, stuck users, slow steps: ranked by users affected, with log lines; the worst one already has a test.
Prompt · Claude Code · Review ·
Help me plan logging for this app so we can fully replay one user's experience: what they did, the key steps inside the app, and what they saw in the end. Give every session an ID, and never log passwords, tokens or personal data. Store the logs where an agent running on a daily schedule can read them. Show me the plan first, and don't write any code until I…
Your agent’s code hides bugs and its tests fool you: a skill it follows on every change, with crapper, dryer and mutator checks.
Skill · Claude Code, Cursor, Grok · Codebase ·
Install the ub-mode skill from unclebob/ub-mode on GitHub into my skills folder, and use /ub-mode for my next change. If crapper, dryer or mutator aren't installed, skip those steps.
A landing page that looks designed: layout, copy, fonts and spacing handled after one round of questions.
Skill · Claude Code, Codex · Design ·
Install the landing-page-design skill from the elayadesign/ai-design-skills repo on GitHub into this project's .claude/skills, then use it to build a landing page for this project. Ask me all your questions in one go before you start.
Keep hitting the usage limit: 7 days of session logs scanned, the 3 costliest habits and their share, rules to plug them.
Rules · Claude Code, Codex · Models and cost ·
Change nothing until I say yes. Write a small script to scan my Claude Code session logs in ~/.claude/projects from the last 7 days, using the token counts recorded in them; don't read the raw files into context. Find the 3 habits that burned the most usage, like reading whole files, dumping long test or build output, spawning subagents, or keeping one sessi…
The agent codes for races that never happen: one line makes it state the risk and the odds, then wait for your OK.
Rules · Claude Code, Codex · Codebase ·
Before you add code for a race condition or security edge case that is unlikely or purely hypothetical, tell me exactly what would go wrong, how likely that is, and how much complexity the extra code adds, in words a 10-year-old could understand. Then wait for my OK. Don't quietly widen the scope to guard against problems that only exist in theory.
A shot needing old film or old records: the Internet Archive’s free API searched, public-domain items only, each one logged.
Rules · Claude Code · Design ·
When you need old film clips or old record music, read http://archive.org/developers/ first, then search http://archive.org/advancedsearch.php, adding collection:prelinger for films and collection:78rpm for music. Only use items whose licenseurl field contains publicdomain, and skip any item without a licenseurl. Download to media/, trim to the part you use,…
A product video like the viral ones: the 5 of 703 prompts that fit your product, your pick remade as one HTML file.
Prompt · Claude Code · Design ·
Read data/videos.json in https://github.com/yihui-dev/awesome-opus5-5-videos and use only the entries where prompt_partial is false. First read this project's README to understand the product, then pick the 5 entries from motion, explainer, 3d or interactive that would best introduce it, and list their original post links so I can watch them first. Once I ch…
Disk almost full: a read-only scan, a treemap of what development left behind, and a cleanup script only for what you tick.
Prompt · Claude Code · Codebase ·
Scan my home folder read-only. Look at sizes only and don't open any file contents. Skip iCloud, the Photos library and external drives, and skip any folder that triggers a system permission prompt. Find the big folders left behind by development work and build a local HTML page: a treemap sized by space, split into 'safe to remove' and 'let me review first'…
Motion effects you pick by clicking: 120 labelled samples in 5 packs; tick the ones you like and Claude Code builds them in.
Skill · Claude Code · Design ·
Read https://github.com/showheyohtaki/motion-prompt and install the motion-prompt skill in this environment. When it is installed, tell me how to use it in three lines.
Pages that look AI-made at a glance: every tell at desktop, tablet and phone widths, FAIL before WARN, code untouched.
Skill · Claude Code, Codex · Design + Review ·
Install taste-skill by following the README at https://github.com/Leonxlnx/taste-skill (the install name is design-taste-frontend), then read its skill file and scripts.
Then run this project locally and run inspect.mjs once each on the homepage and the main landing pages: inspect.mjs <local URL> --out .taste-skill/inspect/<page name>. Look at the screenshots…
Opus 5.5 videos that look like a slideshow: six camera moves by name with what each one does, plus seconds and speed to add.
Prompt · Any agent · Design ·
→ dolly zoom: the person stays the same size while the background stretches away or closes in
→ rack focus: focus shifts from the foreground to the background
→ FPV drone shot: the camera rushes forward like a drone
What you said you’d do later and forgot: past sessions searched, items ranked, each one a card that opens its own session.
Prompt · Claude Code · Long runs + Codebase ·
Then go through my last 20 sessions (the .jsonl logs under ~/.claude/projects/, not counting this one; the files are big, so search them by keyword and don't read them whole) and find three kinds of work I said I'd do and never did:
1. Things I said I'd "get to later", "deal with next time" or "make a note of".
2. Probl…
Plan written, no code yet: one message for first principles, adversarial review, ablation, Occam’s razor and loose coupling.
Prompt · Any agent · Review ·
Re-analyze this from first principles and run an adversarial review of the current plan. Core mechanisms must be validated with ablation tests. Follow Occam's razor in the design, and keep the code highly cohesive and loosely coupled.
Did the skill edit help: 10 real tasks, 8 yes/no questions, a blind judge subagent; keep only a fix that raises the score.
Prompt · Claude Code · Review ·
Run an eval on [skill name], then improve it:
1. Pick 10 real tasks that used it from the session history in ~/.claude/projects, and save the inputs as-is in evals/. If you can't find enough, ask me for more.
2. Write 8 yes/no questions, each checking one thing: 1 if it's done, 0 if it…
A vibe-coded site that loads slowly: images, fonts, frontend, database, APIs and servers, fixed in one pass and re-measured.
Prompt · Claude Code · Launch + Review ·
Before you ship a vibe-coded site, have Opus 5.5 send out 4 subagents to check these 20 things that slow it down, then fix them all in one pass:
Images and fonts
- Convert every image to WebP or AVIF, sized to how big it actually shows on the page
Making content solo: skills for topics, writing, visuals, video, sound and scheduling, installed and tried on one topic.
Skill · Claude Code · Writing ·
1. last30days · 64K stars
Run it before you pick a topic. It takes the last 30 days of discussion about a topic on Reddit, X, YouTube, HN and TikTok, ranks it by real engagement and writes it up as a brief. Ask /last30days what's exploding in AI agents? and you get…
Opus 5.5 can’t hear its own soundtrack: beats from librosa, the grid checked below 120 Hz, loudness measured on the MP4.
Prompt · Claude Code · Design ·
Add sound to this video. You can't hear, so measure every step and don't go by feel. Install ffmpeg and librosa first if they're missing.
1. Use licensed music: a track I give you, or one under Mixkit's free license. Note the source and the license. Don't synthesize the music or the sound effects in code.
2. Find the beats with librosa. It sometimes locks onto…
Trouble behind the page: keys, billing abuse, payments, monitoring and launch day fixed in one pass; what needs you listed last.
Prompt · Claude Code · Launch + Review ·
Before you ship a vibe-coded site, have Opus 5.5 send out 5 subagents to check these 25 things behind the pages, then fix them all in one pass:
Keys and access
- Keep secret keys like your OpenAI key and Stripe secret key on the server only. Search your fro…
Which job is worth many agents: repo-wide checks, migrations or fix loops, planned on one directory with at most 10 agents.
Prompt · Claude Code · Long runs + Codebase ·
We'll run this as a dynamic workflow in Claude Code. From the second doc, only borrow how it splits and reviews the work.
Then go through this project's code, issues and recent commits, and find three kinds of work that fit a dynamic workflow:
1. Checks that need to cover the whole repo, like hunting for bugs or finding missing permission checks.
Users from Google search: 3 free tools people already search for, your pick built as an SEO page with one button to your app.
Prompt · Claude Code · Launch ·
Read this project and come up with 3 free tools for me to pick from. Each one should relate to my product and be something people search for. Give me the search term for each. Skip broad terms and pick specific ones with little competition. Search each term online and check whether big sites hold the top results. If you can't find the search volume, say it's…
Hover stuck on phones, forms red before anyone types: 47 modern CSS techniques and a list of what your styles get wrong.
Skill · Claude Code, Codex · Design + Launch ·
Install good-css following the README at https://github.com/vojtaholik/good-css, then read its SKILL.md and the reference files you need.
Then use it to check this project's existing styles (CSS files and Tailwind classes in components both count) for three kinds of problems:
1. Things that break on phones: hover states that stick after a tap, and tap areas sm…
What in your project is worth animating: storyboards by Anthropic’s Motion guide, your pick built in HyperFrames.
Prompt · Claude Code · Design ·
Then follow Claude Motion's approach and go through this project's README, landing page, onboarding flow, recent release notes and any charts in it. Find three kinds of things worth turning into an animation:
1. A flow one screenshot can't explain, like the steps a new user takes to get started.
2. A change that falls flat as a…
‘The last 24 hours’ in a scheduled task skips or repeats work: fixed windows, silent read failures and risky posting flagged.
Prompt · Claude Code · Long runs + Review ·
Then review all my scheduled agent tasks against it: cron, launchd, scheduled GitHub Actions workflows, and scripts that call claude -p or codex exec. If any run on a server or in the cloud where this machine can't see them, tell me which config to paste in. Find three kinds of problems:
1. Tasks that read a fixed time window, like 'the last 24 hours' or 'sin…
Background you re-explain every new session: pulled from your last 20 sessions and written up as project instructions.
Rules · Claude Code · Codebase + Models and cost ·
Then check my last 20 sessions (the .jsonl files in this repo's folder under ~/.claude/projects/, only what I typed myself) and my recent commits against it, and find three kinds of things:
1. Work that keeps producing new tasks: bugs fixed again and again in the same code, or migrations and features that took several sessions to finish. Using the doc's 'When…
Zero sales and you can still be sued: under-13s, Google fonts, session replay, marketing email, renewal terms, DMCA, all fixed.
Helper agents · Claude Code · Launch + Review ·
Claude Code tip: a vibe-coded app can get you sued with zero sales. Before you ship, have Opus 5.5 send out 6 subagents to check these 6 legal traps, then fix them all in one pass:
No under-13 check at signup
→ Collecting personal info from kids under 13 without parent…
AGENTS.md past 200 lines gets followed less: 3 slimmer versions, each more aggressive, nothing lost, shown before changes.
Prompt · Claude Code, Codex · Codebase + Models and cost ·
Read these two docs, then read my AGENTS.md (and CLAUDE.md, if the project has one) and the files it imports, and give me three ways to restructure them:
1. Delete instructions that do nothing: things Opus already does, and things it can tell from the code. Merge duplicates, and list any contradictions so I can choose.
2. Keep only the rules needed every time…
The cache expires without warning: the status line shows minutes left, then what the next message re-caches.
Settings · Claude Code · Models and cost ·
Read this doc, plus the status line field reference it links to, and add a cache segment to my Claude Code status line.
1. If I already have a status line, don't replace it. Save stdin first so the original command and the cache segment read the same input, leave the original output as is, and put the cache segment on its own line after it. If I don't have o…
What it needs from you is buried mid-report: one AGENTS.md section puts the to-dos, numbered, at the top of every report.
Rules · Claude Code, Codex · Long runs + Codebase ·
When a long task is done, or you're stuck and need me to do something, list what you need from me at the very top of your reply. One thing per item, numbered. Spell out what I should do, where, and what to send back to you when I'm done, e.g. "Log into Stripe, export the September refunds CSV, put it in inbox/". Whatever's blocking you most goes first. If yo…
A week of sessions you can’t recall: local logs as an HTML calendar, no-commit sessions flagged, and a 3-line weekly report.
Helper agents · Claude Code · Long runs + Writing ·
Set up a subagent that only builds weekly calendars:
1. Create week-calendar in ~/.claude/agents: model: opus, effort: medium, memory: user. Preload the design skills I have installed (for example impeccable). Give it Bash so it can run git log and the script that builds the calendar. It only…
The agent keeps making one mistake: the right way in a helper, the old way banned by lint, an error that names the helper to use.
Rules · Claude Code, Codex · Codebase + Review ·
Go through the corrections in AGENTS.md or CLAUDE.md and the last 50 commits in this repo (focus on fixes and reverts). Find patterns the agent keeps getting wrong that lint can catch, up to 3. If you find fewer, report fewer. For each one:
1. Write a helper function or type that wraps the right way to do it
A fast browser for Claude Code: ChatGPT’s computer use as a local MCP server, clicking pages and comparing live with local.
Connector · Claude Code, Codex · Design + Review ·
Set up Codex's computer use as an MCP server for you on my Mac:
1. If there's already an MCP server called codex-cu and it works, skip to step 3.
2. Otherwise, open the .mcp.json in the newest version folder under ~/.codex/plugins/cache/openai-bundled/unified-comput…
A quality bar for the site it builds: each round screenshotted, scored on the Awwwards sheet, until every category hits 7.5.
Prompt · Claude Code · Design + Review ·
Quality bar: score yourself on the Awwwards scoring sheet. Design 40%, usability 30%, creativity 20%, content 10%, each out of 10, as strictly as a real jury would. After every round of changes, run the page and take a screenshot first, then score while looking at it. Write down what's holding back the lowest category, and keep going until every category is…
Pages look AI-made since Opus hasn’t seen what you like: 3-5 screenshots become a DESIGN.md; add taste-skill and Impeccable.
Rules · Claude Code · Design ·
These are screenshots of sites I like. Don't touch any code yet:
1. For each one, tell me what it does well: layout, typography, color, spacing, white space.
2. List what they have in common, ask me which parts I want, and wait for my answer before going on.
Two agents, one rules file: plan first, smallest change, subagent roles, own the bug, verify before done, Lessons per correction.
Rules · Claude Code, Codex · Codebase + Long runs ·
Shared by Codex and Claude Code. Every line here should change what you do.
## 1. Plan before you touch code
- Long task: first tell me in 2-3 sentences what you think I'm after
Today’s pushes to main, reviewed while you sleep: a draft PR per bug with a test that fails first, in the cloud, laptop closed.
Command · Claude Code · Review ·
Run every day at 1:07 AM on Opus 5.5 with no connectors attached. Save the instructions below as the prompt, word for word.
First install dependencies and run the full test suite. Note which tests were already failing. If the tests can't run, or there are none, write down what's missing and stop.
Then look at the commits pushed to the default branch in the las…
Astra eats a week’s quota: 6.1 Sol on high as orchestrator, on medium for 3 subagents, Astra only to review big changes.
Helper agents · Codex · Models and cost ·
Set up my Codex to match this tree:
→ Main session: default to `gpt-6.1-sol` on high reasoning effort. It breaks down tasks, delegates, integrates, and verifies. Handle simple tasks directly; don't delegate them.
→ Three project-level custom agents, each with model `gpt-6.1-so…
No memory of code, notes, sessions or long tasks: 8 repos in those 4 kinds, one of each installed on your OK.
Connector · Claude Code · Codebase + Long runs ·
Help me build a second brain from these 8 GitHub repos: DeusData/codebase-memory-mcp, Graphify-Labs/graphify, garrytan/gbrain, AgriciDaniel/claude-obsidian, tobi/qmd, thedotmack/claude-mem, mem0ai/mem0, gastownhall/beads. Pick by four categories: code graph, knowledge base, session memory, task progress, at most one per category. First check what CLAUDE.md a…
Your costliest Claude API feature: a test set first, then cheaper models and lower effort step by step while accuracy holds.
Skill · Claude Code · Models and cost ·
Then list every place in this project that calls the Claude API, estimate which one costs the most, and let me pick one. Use build-eval from the claude-api skill to build a test set for it, and tell me roughly what it will cost in API fees before you run anything.
Once I've approved the test cases and the grading, run hillclimb on a new branch. The goal is to…
Stuck on style, components, motion, a demo video or polish: 8 sites sorted by that, saved so Claude Code knows which to use.
Rules · Claude Code · Design ·
Add a section called Frontend references to ~/.claude/CLAUDE.md. Use it only when building a new page, when I say something looks bad, or when I name one of these sites. For small changes, just do the work:
- Style: pick a DESIGN.md that fits the product from styles.refero.design or VoltAgent/awesome-design-md on GitHub. Put it in the project root and add an…
A motion video in Claude Code: breakdown, storyboard and review subagents around a main session on HyperFrames; you direct.
Helper agents · Claude Code · Design ·
Set up a motion video team in ~/.claude/agents that I can use for every product video. If HyperFrames isn't installed, install it first following hyperframes.dev.
1. reference-breaker (effort: medium): use ffmpeg to pull one frame per second from the reference video I give you, tile them into one image to look at, and break down the pacing, type, colors and t…
Try to break it yourself first: 16 questions on security, data, flows and environment, issues ranked by severity, then fixed.
Prompt · Claude Code · Launch + Review ·
Before you ship, have Opus 5.5 send out 4 subagents whose only job is to break your app, and split these 16 questions between them:
Security
- Do your bundled frontend JS or git history contain private keys like sk_ or service_role?
Where a long task stands while you wait: an HTML dashboard of progress, blockers and open questions, built by a side subagent.
Helper agents · Claude Code · Long runs ·
Set up a subagent that only builds progress dashboards:
1. Create dashboard-builder in ~/.claude/agents: model: opus, effort: medium, memory: user. Preload the design skills I have installed (for example impeccable). It can only read and write the .dashboard/ folder and its own memory.
2. The fir…
A vibe-coded site about to go live: design, mobile, every state, real-user flows and launch details, fixed in one pass.
Prompt · Claude Code · Launch + Design ·
Before you ship a vibe-coded site, have Opus 5.5 send out 5 subagents to check these 20 things, then fix them all in one pass:
Consistent design
- Write a DESIGN.md first, and make every color, font size, spacing and corner radius follow it
Image options drawn one at a time burn your limits: a global rule puts them all in one numbered grid, then builds your pick.
Rules · Codex · Design + Models and cost ·
Add an image rule to my global Codex AGENTS.md (in ~/.codex; if an AGENTS.override.md exists, add it to that file):
- If it's unclear what I want (what it is, who it's for, what style), ask me one or two questions first. If it's clear, just generate the image.
- When I ask for multiple options, variants or style comparisons, put them all in one image: a g…
Where your setup wastes tokens each turn: measured, each instruction line marked Keep, Move or Cut, applied on your OK.
Prompt · Claude Code · Models and cost ·
Opus 5.5 is already priced low per token. Find what in this Claude Code setup still wastes tokens on every turn. Goal: the lowest cost per finished task. It must stay as thorough as it is now and keep taking on big tasks. Smaller requests don't count as a win if the task then takes more turns.
1. Measure first, change nothing
High effort mostly buys verification: a builder on low, a verifier on high, one CLAUDE.md rule runs each feature through both.
Helper agents · Claude Code · Launch + Review ·
Set up the workflow in this image in my Claude Code:
1. Create two subagents in ~/.claude/agents:
- builder: writes the code. model: opus, effort: low. It implements the spec I've approved and stops when it's done.
Settings carried over from Opus 5 cost more on 5.5: effort, subagents, MCP servers and CLAUDE.md checked, a small fix each.
Prompt · Claude Code · Models and cost ·
Then review my Claude Code setup against it: my user and project settings, the subagents in ~/.claude/agents and .claude/agents, CLAUDE.md, and the MCP servers I have connected. Find three kinds of problems:
1. A default effort that doesn't fit the work. Use the guide's ladder: low for mechanical edits, medium for well-scoped daily work, high when medium stal…
A cleanup that could delete your chats: items sorted Keep, History, Work, Clearable or Unknown, each command approved by you.
Prompt · Codex · Codebase ·
Disk audit: ~/.codex and ~/.claude
Report first. Change nothing until I approve.
1. Measure: du -sh both folders, the 20 biggest items in each, and Codex sessions/ + archived_sessions/ by month. Sizes,…
Fable quota idle while Opus 5.5 does the work: Fable called in before a plan, when an error repeats and before ‘done’.
Helper agents · Claude Code · Models and cost + Review ·
Set up my Claude Code to match this tree:
1. For the three subagents, reuse fitting ones from ~/.claude/agents and .claude/agents, and propose new ones only for missing roles. Set each to model: opus and effort: medium. Leave any subagent that already sets a different model alone,…
Opus 5.5 already thinks before it answers: rules that push extra thinking, dole out steps or stop to confirm, quoted with a fix.
Rules · Claude Code · Models and cost + Codebase ·
Read this guide and review my project's CLAUDE.md, AGENTS.md, skills, and saved prompts. Find three kinds of problems: lines that push you to think more, like 'think carefully' or 'take a deep breath'; rules and prompts that hand out tasks one step at a time or don't define what done looks like; and rules that make you stop for confirmation when you don't ne…
A task longer than one session: progress, decisions, files, checks and next steps kept current for the next agent.
Prompt · Claude Code, Codex · Long runs + Writing ·
Implement [REQUIREMENT] and keep HANDOFF.md updated with current progress, key decisions and why you made them, relevant files, verification results, unresolved issues, and next steps. Keep both DX (developer experience) and AX (agent experience) in mind. Retain only information that’s still current, so I can understand and review the work, and the next agen…
Still walking a senior AI through every step: tech lead, requirements, debugging, performance, frontend, backend, reviewer.
Prompt · Claude Code, Codex · Long runs + Review ·
I want to complete this: [TASK OR AGREED PLAN].
Pick up from the current state and keep going until [THE FINAL RESULT I CAN SEE, USE, OR CHECK].
Read the relevant files to understand the project. Look up information you need. Use the tools available to you.
Before any app code changes: default, loading, empty, error and success on separate artboards, from a design reference repo.
Prompt · Claude Code · Design ·
Create static UI mockups for [page or feature]. First, read [requirements files and relevant pages] in the current project, and follow its existing design system and components. Then browse the reference repository below, choose a suitable DESIGN.md, and read it in full. Use it to inform layout, typography, and visual styling that the project hasn’t already…
Projects that keep moving while you’re out: OpenAI’s Remote guide applied, what runs alone and what you decide by phone.
Prompt · Codex · Long runs ·
Read this Remote guide and help me design a remote workflow around my projects and work habits. Use the context you already have to understand what I’m working on, then ask about anything that would affect the plan: how long I’m usually away from my computer, how often I can check my phone, and whether my computer can stay awake and online. Don’t give me a g…
Plans made blind to your code and logs: connected to ChatGPT, GPT-6 Pro plans from real PRs and records, then briefs Codex.
Prompt · Codex · Models and cost + Codebase ·
I want to solve [specific problem]. Use the connected [GitHub repository and branch / logs / database] to read the relevant code, recent PRs, and records. Start your response with a brief list of the files and sources you actually read. If you can’t access or verify the requested branch, say so. Analyze the problem against [my goals and constraints], propose…
Something in Claude Code’s interface you want changed: a prompt that finds an existing Mod or proposes the smallest one.
Plugin · Claude Code · Codebase ·
I want to change this behavior in Claude Code: [in the desktop app, hide sensitive values by default and reveal them on hover].
Read the Claude Mods documentation below and check whether my current version and interface support this. First look for an existing Mod or setting that can do it. If it needs a custom Mod, propose the smallest implementation and how…
Rules trimmed for Astra can drop checks Sol and Luna still need: Astra finds what was lost, each model’s rules kept apart.
Rules · Codex · Models and cost + Codebase ·
Review my project’s AGENTS.md and Skills. I use these models: [list your models]. Check whether changes made to streamline instructions for Astra have removed necessary test commands, acceptance criteria, or workflow information, or relaxed necessary confirmation requirements. For changes already made, compare against the previous version. State where eviden…
Rules from older models make Astra over-read or stop early: the triggers, required reading, checks and stop rules to cut.
Rules · Codex · Models and cost + Codebase ·
Read this guide and review my project’s AGENTS.md and Skills. Identify overly broad skill triggers, documents you’re required to read for every task, redundant checks, conflicting instructions, and rules that make you stop for confirmation before the task is complete.
For each issue, quote the original instruction, explain its potential impact, and suggest th…
Before a task goes to Codex: GPT-6 Pro finds what you haven’t thought through and writes a brief an engineer could take.
Prompt · Codex · Writing + Codebase ·
I want to do this: [DESCRIBE WHAT YOU WANT TO DO IN ONE SENTENCE].
Assume an engineer who can solve problems independently will take over, but they haven’t seen our conversation. Using the materials I’ve provided, first help me clarify what problem I’m trying to solve, whether any requirements conflict, an…
Code too complicated, or busy a long time with no clear direction: one finds what needn’t exist, one rechecks the direction.
Prompt · Codex · Review ·
Review this implementation from first principles. Start with the goal and the constraints it must satisfy. Don’t assume the existing design is correct.
Which parts are unnecessarily complex? Which decisions rely on untested assumptions? What can be removed without losing required behavior, and what could be simplified afterward?
List up to 5 changes worth maki…
A reviewer keeps flagging the same things: 50 recent PRs scanned, issues ranked by count, your picks written into AGENTS.md.
Rules · Codex · Codebase + Review ·
Use gh to scan the 50 most recent PRs in each GitHub repo under [CODE DIRECTORY]. Find the issues [REVIEWER GITHUB USERNAME] keeps flagging. Rank them by frequency and link the relevant PRs. Wait for my approval, then add only the rules I choose to ~/.codex/AGENTS.md.
A finish you’d pay a photographer for: headshot, cinematic still, travel, magazine shoot, old-photo repair, golden hour.
Prompt · Any agent · Design ·
Turn this photo into a clean professional headshot for LinkedIn or a company profile. Keep my exact face, facial proportions, skin tone, hairstyle, hairline, age, expression, head angle, shoulder position, body proportions, and recognizable features unchanged. Use a chest-up crop, the same eye-level viewpoint, soft believable studio lighting, and a simple [B…
Asked to change a config, it builds a framework: minimum sufficient change, reuse first, ask before widening, test narrowly.
Rules · Claude Code, Codex · Codebase + Review ·
Complete the current task with the minimum sufficient change.
## Before editing
- Read the relevant code, tests, and configuration directly. Do not work from search snippets or guesses.
Sample data proves little: the real component rendered in every state that matters, fixed only where a failure reproduces.
Prompt · Codex · Design + Launch ·
/goal Check [COMPONENT NAME OR PATH] and make sure it survives the content and states it will encounter in real use.
INSPECT ONLY THIS COMPONENT.
Do not modify it yet. Read its actual props, slots, states, data sources, and where it is used. Based on what it really…
Quota unused before it resets: a UI/UX audit, a clickable code map, measurable speedups, a workspace and a Codex cleanup.
Skill · Codex · Models and cost + Long runs ·
/goal Launch the real site for this project and use the three skills below to audit and improve it:
→ `$ux-audit`: click, type, and complete the core flows like a real user. Find anything confusing, hard to use, or simply broken.
→ `$impeccable`: review the visual hierarchy, typography, spacing, component consistency, mobile experience, an…
Which work deserves its own Bot: six roles with ownership, handoffs and limits, and what stays a Skill or a routine.
Settings · Grok · Long runs ·
This is the Bot I use most. It reads today's tasks and each Bot's checkpoint, sets priorities, then hands the work to the right role.
It also runs the nightly review, but only updates one TEAM_ALIGNMENT.md file. Conflicts, external messages, spending, and changes to important files come back to me.
2. Scout (Research and Leads)
Unused quota overnight: an hour before bed, 3-5 reversible, verifiable tasks from today, prompts ready, nothing changed.
Prompt · Claude Code, Codex · Long runs + Models and cost ·
Run this review every day, one hour before bedtime.
Inspect only sources you already have access to:
→ Repositories changed today, including git status
Pages that could fit any startup: a Refero style as reference, a design plan you approve, then DESIGN.md and screenshot checks.
Prompt · Claude Code, Codex · Design ·
Use <REFERO_STYLE_URL> as the design reference for this project. Do not copy its brand, wording, or full-page layout.
Before writing code, give me a design plan:
→ the page’s audience and single goal
Who finished what across several CLI agents: a goal each, read-only reviewers, worktrees for writers, one owner integrates.
Prompt · Any agent · Long runs ·
Use Orca for [TASK]. One clear goal per worker. Reviewers read only; writers get worktrees. Run in parallel; wait for worker_done. Another agent checks commits.
One owner integrates and tests. No push or prod unless approved.
Report results, blockers and leftovers.
Audit every persistent configuration in this workspace that shapes how you work: project instructions, skills, MCPs, and memory.
Find anything redundant, outdated, conflicting, or irrelevant to this project. First give me a KEEP / REMOVE list with one sentence per item. Do not edit any files.
Wait for my approval. Remove only approved REMOVE items. Keep every…
Relaying conclusions between windows by hand: the task split, a worktree and outcome per agent, one executes, one reviews.
Prompt · Claude Code, Codex · Long runs ·
Run this task with multiple agents in parallel:
[TASK]
You are the coordinator. Check which agents and panes are available, then split the task only where pieces can move inde…
Before you change a repo: click any module for its callers, dependencies, tests and source evidence; a rule keeps it current.
Rules · Codex · Codebase ·
Analyze the current codebase. Do not modify product code. Only create or update files under docs/codemap/.
Ignore vendor, build, dist, cache, and other generated directories.
If docs/codemap/ already exists, first compare the existing codemap.lock with the current repo and list the modules that…
Before it implements on a guess: repo read, goal restated, assumptions and plan listed, then a stop for your approval.
Prompt · Codex · Codebase + Review ·
Work like a contractor who pays for rework. Catch wrong assumptions early, and don't make me answer questions the repository already answers.
1. Investigate before asking
Read the relevant code, tests, configs, dependency manifests, and documentation first. Search the r…
One preference typed in three places: a goal wrapper, plus prompts for components, parallel runs, review, states and corrections.
Prompt · Claude Code, Codex · Long runs + Review ·
task: {your task}. write yourself a full end-to-end goal first, then keep going until the architecture and the result both meet the bar.
split: break that goal into independent pieces and run in parallel only as far as the work genuinely allows. every worker's goal carries its own deliverable, how to verify it, what counts as done, and which files it may touc…
Is your setup as cheap as OpenAI’s harness: pass or fail on each of its 5 rules, with the file and the one thing to change.
Prompt · Claude Code, Codex · Models and cost ·
read through this project's agent config. that includes mcp settings, skills, plugins, hooks and rules files. judge it against these five, one at a time. keep the results separate:
1. how many tools and skills load at the start, and which of those have never been called in this project
2. is there a cap on tool output. which tool has returned th…
Banning AI words one by one never ends: 6 rules from 1946 for your global rules file, plus prompts for docs, PRs and reports.
Rules · Claude Code, Codex · Writing ·
Writing rules, from Orwell, 1946. These govern prose: docs, PR text, messages. Never touch code or technical terms; swap in everyday words only where precision survives.
1. Never use a metaphor, simile or other figure of speech which you are used to seeing in print.
2. Never use a long word where a short one will do.
A skill written from memory drifts from how you work: compared with your past sessions, a dry-run diff, a recurring review.
Skill · Codex · Review ·
read [skill name, workflow file, or folder path] and run a backtest against my past sessions related to it.
prefer sessions from the same project and same task type. first list the samples you plan to use and why you picked them. if you can't access enough sessions, say so directly instead of guessing.
compare what the skill assumes vs how i actua…
Many agents open and the Mac bogs down: a dry-run map of processes and startup items, cleaned only where you approve.
Prompt · Codex · Long runs + Review ·
Run a coding-agent performance audit on this computer. Detect macOS or Windows and record the hardware configuration first. Use only built-in, read-only tools. Do not spawn subagents or modify anything on the first pass.
1. Establish a baseline.
Take several samples…
‘Fix this bug’ hands you defaults: what your diff touches, an end-to-end goal, a red-teamed plan, a user walk, a thinking budget.
Prompt · Claude Code, Codex · Review + Codebase ·
i'm about to change [file/module]. don't write any code yet.
use codebase memory to map the repo into one graph, read my uncommitted diff, map every change to the functions it touches, and return: each function, its downstream blast radius, and a risk level.
once it's ranked by risk, act at three leve…
A review with a designer’s authority: taste baseline, correctness fixed, cheap-looking parts found, big changes proposed.
Prompt · Claude Code · Design + Review ·
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 charac…
Dozens of grep and read turns per question: the repo as one graph of functions, files and dependencies; dead code found too.
Connector · Claude Code, Codex · Codebase ·
Set up codebase-memory on this repo for me and show me the map.
Install it from https://github.com/DeusData/codebase-memory-mcp:
curl -fsSL https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/install.sh | bash -s -- --ui
Fable 5 burns tokens fast: planning and frontend stay with it, the backend goes to Codex by /goal; a redesign prompt too.
Prompt · Claude Code, Codex · Models and cost ·
to save tokens, keep this main session (fable 5) on planning and frontend tasks, its visual output and ideas are worth the price. for backend and heavier implementation, write a clear spec and dispatch to codex (gpt-5.5 xhigh) with /goal to execute, my quota there sits unused anyway. you may keep the hardest parts in this session.
The agent over-designs and edits what it wasn’t asked to: think before coding, simplicity first, surgical edits, goal-driven work.
Rules · Claude Code · Codebase ·
- think before coding: ask when unsure, don't silently pick one interpretation and run with it
- simplicity first: minimum code, any overengineering shows at a glance
- surgical edits: only touch what's required, don't fix up neighboring code on the way by
Someone put 314 malicious skills on ClawHub: check the source, scan the code, question the permissions, in under five minutes.
Rules · Any agent · Review ·
1. source: who's the author, do they have other legit projects
2. code scan: search for curl, wget, exec, external URLs
3. permissions: does it need file system access? shell? why?