rjv-verify-skilllisted
Install: claude install-skill rjvim/ai-skills
# Writing a repo's verify skill
A verify skill gives an agent hands and eyes on the running app. It
lives in the repo, next to the code it checks, and every agent runs it
before a commit.
## 1. Place and name it
| Repo | Skill |
|---|---|
| one app | `.claude/skills/verify/SKILL.md` |
| several apps | one per app: `.claude/skills/verify-<app>/SKILL.md` |
Codex reads the same files when the repo's `AGENTS.md` points to them.
An existing local-test skill (for example `carnelian-local-test`) is a
verify skill already: harden it, never add a second one.
## 2. Write five sections
1. **Boot**: the exact command, port, env file, and the ready line to
wait for. Reuse a server that is already running; check the port
first. Seed data with one command.
2. **Drive**: the tool for this kind of app. Web: `agent-browser` with
`rjv-agent-browser-notes`. Mobile: the simulator or emulator MCP.
Backend: `curl` or the API harness. Desktop: computer use.
3. **Feature map**: each user flow as path, selectors or calls, and
the expected result. This is the rubric the agent grades against.
4. **Inspect**: what counts as proof for each flow: a screenshot at a
named width, a response body, a log line, a row in the database.
5. **Teardown**: kill every server, browser and emulator the run
started.
## 3. Make it fast and reliable
It is the smoke harness for every change, so a slow or flaky one gets
skipped.
- Time each section. Boot to ready stays under 2 minutes; one flow