Create a verification skill

Create one compact, project-specific SKILL.md through skill_manage, which shows the person every change with an Undo. Do not install skill files directly or silently add scripts to the user's project.

1. Inspect before drafting

Use the tools already mounted on this bot to establish:

• the exact project or app being verified;
• its existing launch command and reliable ready signal;
• the safest available control surface: existing project CLI, browser, computer, phone, or shell;
• one read-only health check;
• up to three important user workflows and the observable result that proves each one worked;
• how to stop only the process or resource started by the verification run.

Prefer an existing control surface. If the project has no dependable way to launch or drive the relevant workflow, report that prerequisite instead of inventing commands, selectors, APIs, or a new control program.

2. Prove one workflow

Run the real launch, health check, one representative workflow, evidence capture, and cleanup once. Use an isolated test profile or temporary data when available; never drive the user's live data merely to author the skill. If the workflow cannot be proved with the tools currently mounted, stop and explain which capability is missing.

3. Draft one self-contained skill

The proposed SKILL.md must contain YAML frontmatter and these sections:

1. Launch — exact command, isolated data/profile, ready signal.
2. Doctor — one read-only check with an actionable failure message.
3. Drive — stable names, accessibility targets, routes, refs, or commands; never recorded screen coordinates.
4. Evidence — action plus resulting state and any important side effect.
5. Cleanup — stop only what this run started; preserve the evidence.
6. Feature map — at most three proven workflows, each with the user path, control recipe, success proof, and known gotcha.
7. Maintenance — re-run Doctor and the affected workflow when the app changes; update drifted instructions but never hide a product failure.

Keep it concise. Evidence about the user path is input to the draft—not permission to retain secrets, audio, screenshots, or coordinates.

4. Save it

Call skills_list first. If an existing skill already covers the project, report that and do not overwrite it. Otherwise call skill_manage with action="create", the complete skill_md, a short gist, and the exact source used (repository path, URL, or conversation).

Read the skill_manage result. If it applied, tell the user the skill's name, that it is on, and that they can undo it from the chat. Only if it is pending, tell them it stays inactive until they decide on the in-app card. Never claim it was enabled without an applied result.