EvoZeus-Skill Proposal
Skill changes alter agent behavior. Treat them as governance-risk work, not normal docs polish.
When A Skill Is Justified
Propose a Skill when repeated evidence shows:
- agents fail the same workflow under pressure
- the fix requires judgment, not only a script or regex
- the pattern applies beyond one private session
- the public instruction can be written without leaking private context
- success and rejection criteria are clear
Do not create a Skill for a one-off result, generic advice, or behavior that should be enforced by code.
Proposal Shape
Include:
- trigger condition
- failure evidence
- desired behavior
- non-goals
- safety boundary
- related ontology terms
- public/private evidence status
- rollback or deprecation path
Use ../../../.github/PULL_REQUEST_TEMPLATE/skill_instruction_change.md when opening a PR.
Editing Existing Skills
- Keep root
../../../SKILL.mdas the stable session judgment entry. - Put scenario-specific rules under
../. - Keep descriptions as trigger conditions only; do not summarize the whole workflow in frontmatter.
- Reuse project terms from
../../../docs/governance/terminology-glossary.md. - Run
python3 scripts/check_pr_ready.py --base <base-ref>from repository root before review.