AGENT SKILL

targets

暂无简介

License 许可
MIT
最近更新
2026-10-26
Stars
★ 0
Forks
0

02 资源文件 3

03 SKILL.md 正文

Target: a skill

Read when: the target is a skill directory (a SKILL.md with or without references and scripts).

Exposure: --plugin-dir (a loose skill is wrapped in a generated plugin whose name you set with wrapper_name), --add-dir so the worker can read the skill's own files. The prompt never names the skill, with one exception: a skill with disable-model-invocation: true cannot be chosen by a worker, only started by a slash command, so its prompt has to begin with /<wrapper_name>:<skill> (the dry-run tells you the exact form). Such a run tests what happens after the skill starts; whether it would be chosen (T1) cannot be observed for it.

Lens, in order:

  1. Trigger (T1-T3). Was the skill chosen? The listing the worker saw is in env.json; the description is the only thing it had to decide with (plus the name). Several takes: report the rate.
  2. Discovery (S3, S4). Did it open the files the text points to? Which references were never read?
  3. Following (M1-M6, S1, S2). Where did it guess a threshold, a path, a format? Where did two instructions conflict, and which one won?
  4. Execution (O1-O5). Scripts the skill ships: do the documented command lines work from the directory the worker is in?
  5. Completion (C1, C4). Does "done" match the receipts?

Typical scenarios: a normal task in a user's words; an ambiguous one; one with a missing prerequisite. A skill with a references/ directory also needs a scenario that can only be solved by reading one.

To write or redesign the skill under test, use /agentic-maker:build-skill; this page is for testing it.

Do not test by telling the worker "use the X skill".

04 同仓库更多 Skills