write-skill

Files (2)

• SKILL.md
• SKILL_STYLE.md · references/

SKILL.md (first lines)

# Write Skill

A skill is guidance for a capable model, not a tutorial or a workflow engine.
Prefer folding a lesson into an existing owner over creating another component.

- Lead the description with the situation that needs the skill. The body adds
  the stance, failure modes, local standards and boundaries that ordinary
  competence misses. Do not repeat the trigger as a feature tour.
- Use clear principles and observable outcomes. Sequence only when order prevents
  a real failure; require a schema only for a real consumer. Compact metaphors or
  leading words are hypotheses about behavior, not proof that a prompt binds.
- Preserve non-obvious constraints when cutting prose. Each deletion needs a
  reason the behavior is redundant, relocated or deliberately retired. Factor
  shared detail into a reachable reference, keeping the condition for reading it
  precise. Loaded references still cost context; progressive disclosure is not free.
- Default to discoverable skills with narrow triggers. Use
  `disable-model-invocation: true` for genuinely user-driven entrypoints, not a
  budget shortcut. Hidden discovery does not make files unreadable or enforce
  write authority. Gate side effects explicitly and respect native permissions.
- Resolve helpers/references from the loaded skill directory. A selected skill
  must not silently require an uninstalled sibling: use an optional shared owner
  with a focused fallback, not a copied framework or dependency installer.
- State is usually files; native host tools own runtime behavior. A new skill,
  agent, extension or hook must justify a capability the existing owners cannot
  supply. Suggest consequential additions/retirements before making them.

[SKILL_STYLE](references/SKILL_STYLE.md) gives the deeper review criteria. Respect
the repository's current memory convention; no sibling memory skill is required.
After an edit, audit overlap, discovery, references and conditional payload cost.
Separate static checks from authorized consuming-agent evaluation and report
missing behavioral evidence honestly.

Deliver: ownership decision, changed source paths, preserved/retired behavior,
checks and remaining uncertainty. Do not deploy or delete user copies implicitly.

From bryan824/kirin-pi, MIT.

More from bryan824/kirin-pi