跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
浏览器与自动化

create-verify-skill

Create a missing project verify-<app> skill using agent-browser for web apps or tuistory for terminal apps. Use maintain-verify-skill for drift, /create-skill for generic skills, or agent-browser or tuistory for a one-time drive.

Cynnamoroll · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add Cynnamoroll/factory-stack --skill create-verify-skill

技能说明


name: create-verify-skill description: "Create a missing project verify- skill using agent-browser for web apps or tuistory for terminal apps. Use maintain-verify-skill for drift, /create-skill for generic skills, or agent-browser or tuistory for a one-time drive." license: MIT compatibility: droid disable-model-invocation: true

Create verify skill

Generate .factory/skills/verify-<app>/ to launch the real app, drive a user's feature path, and capture evidence. Assume the next agent has never seen the app.

1. Interview the repo

Answer from code. Ask only what you cannot observe.

  • Surface. What does a user touch: a web UI, a CLI or TUI, a desktop app, an API, a library? Pick the primary one and note the rest.
  • Run. The repo's own dev command (package scripts, Makefile, README), ports, env vars, seed data, auth.
  • Drive. Existing harnesses first (Playwright or Cypress specs, PTY helpers, curl-able endpoints). Otherwise use Droid's built-in agent-browser skill for web and Electron apps, tuistory for terminal apps, and plain HTTP for services.
  • Observe. Screenshots, terminal transcripts, response bodies, logs, exit codes, stored state.
  • Isolate. Can two instances run side by side (ports, data dirs, profiles)? If not, the generated skill says so and refuses to drive a shared instance.

If the checkout does not build or start, fix that or report it precisely before generating.

2. Write the skill

Write .factory/skills/verify-<app>/SKILL.md with name: verify-<app> and a description naming the app, surface, and trigger. Fill all sections from the interview, without placeholders:

  • Launch. The exact command, how to tell it is ready to use (not just that a port answers), and teardown. Launch works from an empty data dir, including migrations and seed data. A short-lived CLI builds once, then runs each drive in its own PTY session.
  • Doctor. One read-only check: process up, right build, port owned by this run, auth valid. Run it first whenever anything looks off.
  • Drive. The harness recipe with real selectors and commands from this repo. Prefer stable handles (ARIA names, data attributes, prompt strings, routes) over coordinates.
  • Evidence. What to capture and where. Drive the real user path, not internal setters or test-only endpoints. Capture the action and the resulting state. Check side effects (files, rows, messages) as well as the screen. Verify what a dry-run mode really skips by watching it.
  • Cleanup. Kill only what this run started, never by process name. Remove instances and scratch state. Keep the evidence at the path the skill names.
  • Helpers. Every script the skill ships is executable, and its invocation is shown in the body.

3. Seed the feature map

Create .factory/skills/verify-<app>/features/README.md and one file per user-facing feature, starting with the top 3 to 5 from routes, commands, menus, or docs. Follow the shape in references/feature-map-example.md. If seed data lacks a precondition a journey needs, the journey reports a failure and the reason goes in that feature's Gotchas. Never weaken a journey so it passes.

4. Prove it

Follow the generated skill end to end once: launch, doctor, drive one mapped feature, capture evidence, clean up. Then confirm the evidence still exists at its named path. Fix whatever fails, and run the generated cleanup after every failed attempt. If you proved it in a stand-in (a container instead of the real target), list the setup steps the stand-in did that the target will not.

Done when

  • .factory/skills/verify-<app>/SKILL.md exists with valid frontmatter and all six sections filled.
  • The feature map has a README and at least three feature files.
  • One full run passed, its evidence survived cleanup, and the reply links the evidence path.
  • The reply points the user at maintain-verify-skill for upkeep.