rageclick

Was macht rageclick?

Turn a URL and a user goal into a short, reproducible sequence that breaks a promised behavior. Use the existing browser tool to inspect the real page, then the bundled Playwright runner for repeatable evidence. This skill does not rate visual design or claim to represent real user research.

Scope and success criteria

Work only on owned fixtures, local apps, or explicitly authorized test environments. The bundled runner accepts loopback URLs only, starts a fresh browser context, and blocks requests outside the exact origin. A local app may still mutate a real backend through its server: establish synthetic data and a disposable account before submitting. Do not buy, send messages, delete valuable data, or change access settings. Existing permission covers the agreed workflow only.

From the user's goal and observed UI, identify one concrete invariant: one project for one submission intent, recovery after an invalid file, or a draft surviving a promised save. If the contract is ambiguous, record an observation or ask about that specific behavior instead of inventing a defect. Record the application version or commit; use unknown if unavailable.

Explore, perturb, replay

1. Inspect the current page using an available browser tool or the target source. Select stable labels, roles, or test IDs from actual evidence. Run the ordinary successful workflow first. If it cannot complete, resolve that prerequisite or report it separately.

2. Read references/plans.md to write a JSON plan. Choose the applicable templates from references/perturbations.md. Change one dimension per case. Fresh browser contexts isolate browser state, but server data requires a known reset or a unique disposable account. Do not assume a fresh browser resets a database.

3. Run from this skill directory after npm ci and npx playwright install chromium:

   node scripts/run.mjs --url http://127.0.0.1:3000 --plan /absolute/path/case.json --out /absolute/path/evidence --app-version COMMIT
   

   The output directory must be new. The runner writes a readable report, exact plan, step results, screenshots, trace, video, and replay script. It cannot infer a goal or selectors by itself; the skill does that work. Keep generated evidence outside the target repository unless requested as a durable example.

4. Treat observed_failure as an assertion mismatch, not an established root cause. Repeat the same case in fresh state at least twice before calling it reproducible. Capture failures and passes; never discard contradictory reruns. execution_error means an action, selector, browser, or boundary failed and is not product evidence. Record subjective UX comments separately from objective assertions.

5. Minimize the sequence without changing the failure, then use replay.mjs as a regression command (exit 0 passes, 1 assertion mismatch, 2 execution/configuration error). If authorized to fix the app, rerun the identical plan against the fixed version and verify the ordinary path still passes. Never change the assertion to manufacture a passing result.

Report the violated contract, actual and expected state, occurrence count, app/browser versions, exact commands and evidence paths. Say when video, trace, baseline, reset, or repeatability is missing. Redact credentials and personal data before sharing captures; traces can contain complete request/response data. No publishing or external messages are implied by this skill.

Limits

The runner uses Chromium, CSS selectors, native pointer events, synthetic inline files, and bounded request delays. A request delay is not bandwidth throttling or packet loss. It cannot automate remote staging, HTTP redirects (including same-origin redirects), streaming responses, multi-origin login, service workers, WebSockets, or arbitrary JavaScript. Responses are buffered through Playwright with redirects and transport retries disabled, so streaming and network timing fidelity are limited. Browser credential and Set-Cookie policy testing is also unsupported: API transport can store response cookies differently from browser fetch. For unsupported workflows use an already available, authorized browser tool and preserve the same evidence contract; mark unsupported perturbations explicitly. Do not weaken the runner's origin boundary to make a test pass.