Prove It

Skills that make your coding agent prove its work.

Your agent says "done". The tests pass. The code looks right. Then you open the app and the button does nothing.

Prove It gives your repo its own verification skill: one that launches your real app, drives it the way a user does, saves the evidence and cleans up after itself. Then it keeps that skill honest, and tries to fool it.

Works with Claude Code, Codex and any agent that reads SKILL.md files. Free and MIT licensed.

Learn it in 7 days. The free email course teaches one idea a day, using real repos and the mistakes agents actually make. You'll also hear about new skills when they land.

Get the free 7-day course

Install

Claude Code

/plugin install prove-it --marketplace ray-amjad/skills

Codex

codex plugin marketplace add ray-amjad/skills
codex plugin add prove-it@ray-amjad

Any other agent, or files you can edit

npx skills@latest add ray-amjad/skills

Or copy the folders you want from skills/ into ~/.claude/skills/ (Claude Code) or ~/.agents/skills/ (Codex).

Then, in the repo you want to verify:

/create-prove-it

Why it exists

Coding agents fail in the same few ways. Each skill fixes one.

1. "Done" means the code compiles

The problem. The agent reads its own diff, runs the tests and calls it done. Tests mock the parts that break. Nobody opened the app.

The fix. /create-prove-it gives your repo a prove-<app> skill that runs the real app, so any agent can type /prove-<app> before it says done. It starts by asking where users actually meet your app (the web page, API, CLI, database, emails, background jobs) and whether your agent can reach each of those from your machine. Where it can't, it suggests a way in: a sandbox, a local emulator, a seeded test account or a headless browser. Then it shows you a brainstorm:

Method	Catches	Misses	Cost	Reachable now?
Browser, driven like a user	The open count not updating after a tick	Whether the task was really saved	Chromium once, ~5s a run	Yes
API calls	Empty titles accepted, wrong status codes	What the page does with the response	Free, under 1s	Yes
Stored-state check	"Saved" on screen, nothing on disk	Whether a user can see it	Free	Yes
Email capture	Reminders that never send	Real delivery	A local mail catcher	No: needs `MAIL_API_KEY`

You pick the methods. It writes the skill using your repo's real commands and selectors, then runs it once, end to end, and shows you the receipt.

2. The verifier goes stale

The problem. You rename a button, add a feature, change a port. The verification skill still describes last month's app, and the next agent trusts it.

The fix. /maintain-prove-it checks git for what changed since each feature was last proved, then re-proves those features in the running app. When the app and the skill disagree, it works out which one is wrong. It fixes the skill, and it reports bugs in the app without touching your code.

3. The verifier says PROVED on a broken app

The problem. A check that can't fail is worse than no check, because it gives you confidence you haven't earned.

The fix. /audit-prove-it breaks your app on purpose, one small realistic bug at a time: a counter off by one, a save that never happens, the right output with the wrong exit code. A fresh agent then runs your verifier without knowing what was planted. You get a scorecard, every miss with the fix for it, and an untouched working tree at the end.

In testing, the first audit of a generated verifier planted 7 bugs. The agents doing the checking caught 5, but the verifier's own scripts caught only 1, because every check ran against a single record. That finding now shapes how /create-prove-it writes checks.

The skills

Skill	What it does	When
`/create-prove-it`	Brainstorms how your repo can be verified, then generates a `prove-<app>` skill and proves it works	Once per repo
`/maintain-prove-it`	Re-proves features that changed and fixes the skill where it drifted	After features ship
`/audit-prove-it`	Plants bugs and checks whether the verifier catches them	After creating it, then now and then

What it adds to your repo

your-repo/
  .claude/skills/prove-<app>/      # .agents/skills/ on Codex
    SKILL.md                       # start, health check, drive, try to break it, evidence, clean up
    features.md                    # what your app promises, and how to prove each promise
    scripts/                       # start, health, stop, scripted checks
  .prove-it/                       # git-ignored
    runs/<timestamp>-<slug>/       # receipts, screenshots, logs
    audits/<timestamp>/report.md

License

MIT. See LICENSE.