| name | skill-authoring |
| description | Use when creating, editing, or verifying a skill before deployment. |
Skill Authoring
Treat process documentation as code: create pressure tests, observe failure, teach the minimum behavior, then close demonstrated loopholes. Announce use before authoring.
When to Use
Use for every new skill, behavioral change to an existing skill, invocation redesign, or pre-deployment verification. Do not assume a skill is clear because it reads clearly to its author.
Route
- Before editing behavior, load pressure testing and execute the RED → GREEN → REFACTOR loop.
- When designing metadata, invocation, body organization, pruning, or final quality, load structure and quality.
- Load both only when creating a skill end to end.
Invariants
- RED: run a realistic baseline scenario without the proposed guidance and capture exact decisions and rationalizations.
- GREEN: add the smallest instruction that corrects observed failures, then rerun the same scenario with the skill.
- REFACTOR: add counters only for loopholes actually exposed and rerun all pressure scenarios.
- Combine at least three pressure sources when testing discipline: time, sunk cost, authority, exhaustion, or “pragmatic” exception claims.
- The frontmatter description starts with
Use when and states triggering conditions, not a workflow summary. Choose model- versus user-invocation deliberately.
- Keep most
SKILL.md routers below 500 whitespace words. Move branch-specific detail behind explicit reference links and load only the selected branch.
- Preserve evidence of RED and GREEN; prose review alone is not verification.
Red Flags
- “It is obviously clear,” “testing is overkill,” “we can test later,” or “this is too simple to fail.”
- Rules added for imagined problems, generic advice that changes no behavior, repeated policy, stale sediment, or references loaded unconditionally.
- A description that lets the agent infer the workflow without opening the skill.
- Pressure tests weakened until the existing text passes.
Integration
Use before deploying any skill consumed by workflows, agents, or other skills. Pair with test-driven-development for the discipline and verification-before-completion for fresh final evidence. Feed recurring failures into knowledge-compounding only after the corrected behavior is proven.