uda-eth/proof-skill

proof-skill adds a skill called /proof to Claude Code, the AI coding assistant, so that instead of just saying a task's tests pass, it can show a real user completing the journey the ticket promised. When you run /proof, Claude figures out the user journeys implied by your ticket or code diff, writes a test runner from a bundled template, and drives your actual running app in a real Chrome browser against your real dev server and real database, rather than mocking anything. The README explains why this matters with three real examples of bugs that slipped through despite passing tests: a stale dev server that made every test pass against old code, a feature that worked but was hidden behind an onboarding screen nobody dismissed, and an empty state whose button linked nowhere useful. A journey suite that actually clicks through the app the way a user would catches these kinds of problems that assertions alone miss. Each run produces a folder of evidence next to your pull request: the Playwright test scripts themselves, a machine readable results file, a short markdown report with a pass or fail verdict for each step, and a self contained HTML report page with a screen recording player, before and after screenshots compared against the code before your change, and a five viewport responsiveness sweep from phone to desktop width. The skill enforces a small set of rules, including never mocking the network layer, pairing every screenshot with an assertion, testing negative cases like permissions and filters, and treating anything short of a fully passing suite as not done. Installation is copying the skill folder into your personal or per-project Claude Code skills directory. A demo pomodoro timer app is included to show the full proof pack it generates. It requires Node.js and Playwright, with an optional Postgres database for setting up test users directly, and is released under the MIT license.