I build and run web products, and I write Claude Code skills for the part nobody enjoys: making an agent trustworthy on a system that is already in production.
Every skill here comes from an incident that cost me a real afternoon. Each one lives in its own repository, MIT licensed, so you install only what you want.
| Skill | What it does |
|---|---|
| prove-it | Stops an agent from saying "done" when it never checked. Captures real evidence - HTTP response, browser network request, console errors, screenshot - and returns a PROVEN / NOT PROVEN / UNPROVABLE HERE verdict. |
| safe-worktree | Lets several agents share one repository without stealing each other's work. Runnable isolated worktrees, a pre-commit drift guard, and cleanup that detects a squash merge - which git branch --merged does not. |
| typo-fr | French typography that survives an AI: no em dashes, correct apostrophes, accented capitals, non-breaking spaces. A linter that never touches your code. |
In Claude Code:
/plugin marketplace add JohnJackHouzi/prove-it
/plugin install prove-it@prove-it
From a plain shell:
claude plugin marketplace add JohnJackHouzi/prove-it
claude plugin install prove-it@prove-itSame two lines for safe-worktree and typo-fr. Every script also runs
standalone - no Claude Code, no dependencies beyond git, Python 3 or Node - so
they work in CI or from any other agent.
- One repository per skill. Nothing is bundled you did not ask for.
- No skill ships until its scripts have been run for real, on passing and failing cases. A green build is not evidence that anything works.
- Documentation states what was verified and what was not.
J'écris des skills Claude Code pour ce qui fait mal en production : vérifier vraiment avant d'annoncer qu'une chose marche, travailler à plusieurs sessions sans se voler le boulot, et écrire un français correctement typographié. Chaque skill a son dépôt, sous licence MIT, installable en deux lignes.
Issues and pull requests welcome on any of the repositories above.