<< All versions
Skill v1.0.0
currentAutomated scan100/100leek/agent-skills/verify
──Details
PublishedSeptember 28, 2026 at 02:40 AM
Content Hashsha256:40636cf19ca8544d...
Git SHA
──Files
Files (1 file, 2.1 KB)
SKILL.md2.1 KBactive
SKILL.md · 25 lines · 2.1 KB
version: "1.0.0" name: verify description: Exercise a change end to end in the running application (hit the route, run the command, click through the page) and report what actually happened. Use after tests pass, when a change needs proving in the real app, or when another skill's build step ends.
Verify
Tests passing is not the finish line. Exercise the actual flow once, in the running application, and report what you observed: never what should happen.
Process
- Pick the flow the change owns. The route a user would hit, the command an operator would run, the job the system would dispatch, the Filament/Livewire page a back-office user would open. One realistic pass; this is not a test suite.
- Run it for real.
- HTTP / UI: the app is served (Herd serves Laravel apps at
https://<dir>.test: never start a server that's already running). Use the project's browser tooling (agent-browser, Boost'sbrowser-logs) to load the page, perform the action, and screenshot the result. - Console: run the artisan command with realistic arguments.
- Jobs / listeners: dispatch through the real path (
dispatchSyncor trigger the event), then check the side effect where it lands, database rows, storage, logs.
- Check the seams the tests can't see. Browser console errors, the Laravel log (
storage/logs/or Boost'sread-log-entries/last-error), a queue that needed a running worker, an unbuilt frontend (a Vite manifest error means runnpm run build, not that the change is broken). - Report faithfully. What you did, what you saw (status codes, redirects, rendered state, rows written), and anything that surprised you: with the evidence. If a step couldn't be exercised (needs credentials, a webhook, a human), say so plainly instead of marking it verified.
Guardrails
- Verify against local/dev: never run mutating flows against a real environment's data without explicit approval.
- Leave no residue: throwaway records created during verification get cleaned up, or created under an obviously-fake identity the seeds already use.