AGENT SKILL

create-verification-skill

Create project-local behavior verification skills and adapters when asked to build repeatable UI, CLI, mobile, or service checks.

License 许可
MIT
所属仓库
gutierrezje/jstack
最近更新
2026-11-11
Stars
★ 0
Forks
0

01 前言 Front Matter

name
create-verification-skill
description
Create project-local behavior verification skills and adapters when asked to build repeatable UI, CLI, mobile, or service checks.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

Create verification skill

Read and follow the project verification contract.

  1. Inspect every user-facing surface, its launch path, user journeys, existing tests, shared services, authentication, mutable data, and available control tooling.
  2. Split verification skills when surfaces have materially different control, authentication, state, data-safety, or evidence boundaries. Create one verify-<surface> skill per boundary. Keep one skill only when the same setup, driver, fixtures, cleanup, and proof contract genuinely apply.
  3. Choose the real harness for each surface: $cursor-team-kit:control-ui, $cursor-team-kit:control-cli, browser/chrome/computer-use, or a project-native driver. Reuse an existing environment-setup skill instead of copying its process-management instructions.
  4. Create each project-local .agents/skills/verify-<surface>/ skill with $skill-creator and $writing-for-agents.
  5. Drive the first path manually to learn the real sequence. When repeated project-specific setup, control, observation, or cleanup remains, build the smallest executable adapter that satisfies the shared contract. Test it through its public commands rather than its implementation.
  6. Include exact target identity, setup, launch, a deterministic read-only or disposable happy path, evidence capture, mutation authority, cleanup, failure interpretation, and a feature map under features/. Make launch reach a usable state from an empty data directory, including migrations and seed data. Ready means usable, not only that a process or port answers; a database health check can pass while init scripts still run. When seed data lacks a precondition a journey needs, the journey reports a failure and the reason goes in that feature's Gotchas; never weaken the journey to make it pass. For requested video or temporal behavior, follow the video evidence contract; document recorder capabilities, finalization, receipt coverage, and the recipe's required versus optional video proof.
  7. Seed the top three to five user-facing features in each skill using references/feature-map-example/. Inventory every discovered entry point in the map as covered or explicitly excluded, and name feature IDs so a PR workflow can select only the journeys affected by a change.
  8. Run every adapter or adopted project-native controller against the real product, inspect its receipt, fix the instructions, and validate every generated skill. When the proof ran in a stand-in for the real target (a container for a fresh VM, a local stack for a remote one), list every setup step the stand-in did that the target will not, and keep those steps unverified until a run on the real target covers them.

Completion means a cold agent can select the correct verification skill, bind it to the exact target, discover the control interface through help, drive one real feature per surface, recognize success from a machine-readable receipt, and restore or account for mutated state.

04 同仓库更多 Skills