Each prompt states what done looks like, the check that proves it and a turn bound, so Claude can work through it on its own. Replace the parts in square brackets with your own paths and commands. 10 are free, Pro includes all 32.
frontendFreeruns for about €1.20 of usageabout 25 minSonnet 5.5
Runs an automated accessibility check on key pages and fixes every violation.
/goal the pages [/, /login, /pricing] have zero violations in an automated accessibility check. Start the dev server, then run `npx @axe-core/cli [http://localhost:3000/]` (and the same for each other page) and show the violation count per page. Fix the causes in the components, not per page: missing labels, low contrast, missing alt text, wrong heading order, buttons without names, focus that is not visible. To prove it: run the same checks again and show 0 violations for every page, and run `npm test` and show it exits 0. Constraints: no visual redesign, no aria attributes where a native element does the job. Or stop after 25 turns.
devopsFreeruns for about €0.60 of usageabout 12 minSonnet 5.5
A GitHub Actions workflow that runs the same checks you run locally, with caching.
/goal .github/workflows/ci.yml runs on every pull request and on pushes to main, installs dependencies with the lockfile and caching, and runs this repo's real lint, type check, test and build commands (read them from the project files). To prove it: show the workflow file, then run every command from the workflow locally in the same order and show each exit code, all 0. Constraints: pin actions to a major version, no secrets in the file, set a job timeout, do not push or open a pull request. Or stop after 12 turns.
setupFreeruns for about €0.60 of usageabout 10 minSonnet 5.5
A short CLAUDE.md with the real commands of this repo, each one proven to run.
/goal CLAUDE.md exists at the repo root, is under 120 lines, and lists the install, test, lint and build commands this repo really uses (read them from package.json, Makefile, pyproject.toml or the CI config, never guess). It also lists up to 5 gotchas you found while reading the code, such as required env vars or a service that must run first. To prove it: run every command listed in CLAUDE.md once and show each command with its exit code; any command that needs a missing service or secret is marked in CLAUDE.md as such instead of failing. Constraints: do not change any file except CLAUDE.md, no style rules a linter already enforces, no secrets or real env values. Or stop after 15 turns and show what is still unproven.
testsFreeruns for about €0.80 of usageabout 15 minSonnet 5.5
Fixes the source until a test folder passes, and never edits the tests to cheat.
/goal every test in [test/auth] passes. To prove it: run `npm test -- [test/auth]` and show the summary line with 0 failed. Constraints: fix the source code, not the tests; do not modify, skip or delete any test file or add retries; do not change any file outside [src/auth] and the shared helpers it imports, and list any helper you change with the reason. If a test is wrong, stop and explain why instead of changing it. Or stop after 20 turns and show the remaining failures.
docsFreeruns for about €0.70 of usageabout 15 minSonnet 5.5
Rewrites the README quickstart and proves each command in a fresh clone.
/goal README.md has a quickstart section (prerequisites with versions, install, configure, run, test) that works on a fresh copy of the repo. To prove it: clone the repo into a temporary folder with `git clone . [/tmp/readme-check]`, run every quickstart command in a fresh clone in order, and show each command with its exit code; all exit 0 except steps that need a secret, which the README names and explains how to get. Constraints: only README.md changes (plus an .env.example with placeholder values if the quickstart needs one), no marketing copy, delete the temporary folder at the end. Or stop after 15 turns.
debuggingFreeruns for about €0.90 of usageabout 20 minOpus 5.5
Turns a bug report into a failing test first, then the smallest fix that makes it pass.
/goal the bug [describe the bug and how to trigger it] is fixed and covered by a test. Steps you must show: first add a test that reproduces the bug and show it failing with the same symptom, then make the smallest change that fixes the root cause, then show the same test passing. To prove it: run `npm test` and show the full suite exits 0. Constraints: fix the cause, not the symptom (no special case for the exact input in the report); do not change unrelated files; explain the cause in two sentences in your final message. Or stop after 25 turns.
reviewFreeruns for about €1.20 of usageabout 20 minOpus 5.5
A review pass on the branch diff that fixes real bugs and ignores style opinions.
/goal the changes on this branch have no remaining real bugs. Review `git diff main...HEAD` using REVIEW.md if the repo has one; a real bug is wrong output, a crash, a security hole, data loss or a broken edge case, not style. For each finding show the file, line and a one-sentence failure scenario (input, then wrong result), then fix it and add a test where one is practical. To prove it: run `npm test` and show it exits 0, then list every finding with fixed or not a bug and why. Constraints: no refactors or renames beyond what a fix needs, do not touch files outside the diff unless a fix requires it. Or stop after 20 turns.
securityFreeruns for about €0.70 of usageabout 15 minSonnet 5.5
Finds keys and tokens in the repo, moves them to environment variables and blocks new ones.
/goal no API key, token, password or private key is hard-coded in tracked files. Search with `git grep -nE` for long random-looking strings, private key headers and names like key, secret, token and password with a literal value, and show the search and its results. Move each real secret to an environment variable read at startup with a clear error when missing, add the names with placeholder values to .env.example, and make sure .env is in .gitignore. To prove it: run the same `git grep -nE` search again and show it has no real secrets left, and run `npm test` and show it exits 0. Constraints: never print a secret value in the conversation (show only the file, line and variable name), do not rewrite git history, and list every secret that must be rotated because it was committed. Or stop after 15 turns.
refactorFreeruns for about €1.40 of usageabout 25 minOpus 5.5
Breaks one oversized file into focused modules while every test stays green.
/goal [src/big-file.ts] is split into focused modules under [src/big-file/] where no file is longer than 300 lines, every caller imports from the new modules, and the old file is gone or is a re-export of under 20 lines. To prove it: run `npm test` and the type check (`npx tsc --noEmit`) and show both exit 0, then run `wc -l` on every new file and show the output. Constraints: no behaviour change, no test file is modified except import paths, no new dependencies, public exports keep their names. Commit nothing. Or stop after 30 turns and list what is left.
setupFreeruns for about €0.50 of usageabout 10 minSonnet 5.5
A project skill that runs your tests, type check, lint and build in one go.
/goal .claude/skills/verify/SKILL.md exists with frontmatter (name: verify, a one-sentence description that says to use it before reporting any task as done) and a numbered list of the real commands that prove a change works in this repo: tests, type check, lint, build and, if there is one, a quick smoke run. Read the commands from the project files, never invent them. To prove it: run each step of the verify skill and show its exit code; every step exits 0 on the current main branch, or the step is removed with a one-line reason in your final message. Constraints: SKILL.md under 60 lines, no other files changed except a one-line mention of the skill in CLAUDE.md if CLAUDE.md exists. Or stop after 12 turns.
debuggingProruns for about €1.00 of usageabout 20 minSonnet 5.5