Snippet
Copy and paste this prompt into a coding agent inside your project. It writes a reusable skill that proves the app works, not just that the code compiles:
text
Create a verification skill for this project. It must drive the real app the way a user does, and it must prove behavior with evidence. Unit tests alone do not count.
Read the repository first, and find out:
- Which surface users touch: a web UI, a CLI, a desktop app, an API, or a library.
- The exact command that starts the app locally, plus its ports, environment variables, and seed data.
- Whether a harness already exists, such as Playwright, Cypress, expect scripts, or curl-able endpoints. Use it before you write a new one.
- Whether two instances can run side by side. If they cannot, say so in the skill.
If the project does not build or start, fix that first, or report the exact failure.
Write the skill in the project's skills folder, for example skills/verify-<app>/SKILL.md. Start the file with YAML frontmatter: a name, and a description that names the app, the surface, and when to use the skill. Then include these sections, each filled with real commands and selectors from this repo:
1. Launch: the exact start command, how to tell that the app is ready, and how to stop it.
2. Doctor: one read-only check that says whether the running instance is worth driving.
3. Drive: the steps for each feature, using stable handles such as ARIA roles, labels, or routes.
4. Evidence: what to capture (screenshots, page snapshots, stdout, exit codes, stored data), where to save it, and proof that the real user path was used. Check side effects, not only what is on screen.
5. Cleanup: stop only the processes you started. Remove scratch data, but keep the evidence.
Add a feature map in features/: a README index, and one file per user-facing feature (three to five to start). Each feature file lists its sub-features, the ways a user reaches the feature, the steps to drive it, and its gotchas.
Finally, run the skill end to end once: launch, doctor, drive one feature, capture evidence, and clean up. Confirm that the evidence still exists. Report anything that failed. A skill that has never been run is only a draft.How to Use
- Copy the prompt in the box above.
- Open your coding agent in the root folder of your project.
- Paste the prompt and send it. The agent reads your project before it writes anything.
- When the agent finishes, read the new skill folder. Then ask the agent to run the skill once, and keep the evidence that it saves.
How It Works
- Real user path - The skill drives the app the way a user does, not through test-only shortcuts.
- Evidence you can check - Each run leaves screenshots, snapshots, or transcripts at a known location.
- Repeatable - A feature map lets later runs cover the rest of the app without starting from scratch.
- Proven before handoff - The generated skill must run once before it counts as finished.
Condensed from the create-verification-skill skill in cursor/plugins (pstack, MIT License). The full copy is in skills/create-verification-skill/ in this repository.