One Procedure Per Skill

Keep each skill to a single reproducible procedure so an agent can follow it without branching

Use when: a skill has grown into a survey of several unrelated techniques

Avoid when: the steps genuinely form one sequence that cannot be split

Procedure

1. Read the skill and name the single outcome it produces
2. Split any step that serves a different outcome into its own skill
3. Give each split skill a title that states its outcome
4. Check the cross-references point at the new slugs
5. Confirm the original skill still stands alone
6. Re-run the corpus validator after the split

Signals

• the skill has one stated outcome and one done-when set
• no step could be executed independently of the others
• the title matches the outcome
• the validator passes after restructuring

Patterns

1. Single outcome

produce a cache key verdict, not 'test caching'

2. Split decision

vary inspection separated from poisoning probes

3. Cross-reference

see also: cache-key-header-probe

Pitfalls

• Keeping a catch-all skill because splitting feels like admin work
• Splitting so finely that each skill has three trivial steps
• Leaving dangling references after a split

Done when

• each skill names exactly one outcome
• every split skill passes the validator standalone

References

• OWASP - methodology
• Anthropic - skills