Creating Skills

Create or revise a skill so it predictably changes agent behavior without carrying unnecessary context.

Steps

1. Classify invocation. Decide whether the skill is model-invoked or user-invoked. Keep a trigger-rich description only when the agent or another skill must discover it; otherwise set disable-model-invocation: true and make the description human-facing. Completion: the invocation choice and its reason are stated.
2. Draft the behavior spine. Write the smallest ordered steps that make the skill work. End every step with a checkable completion criterion. Completion: each step says how the agent knows that step is done.
3. Disclose reference. Keep only always-needed steps in SKILL.md; move branch-specific templates, rubrics, examples, command recipes, and long checklists to linked sibling markdown files. Completion: SKILL.md is short enough to scan and every pointer says when to read its target.
4. Prune for predictability. Apply GREAT_SKILL_PRINCIPLES.md: remove duplication, sediment, no-ops, vague encouragement, and weak trigger prose. Prefer one strong leading word over repeated explanation. Completion: every remaining line changes invocation or execution behavior.
5. Add eval hooks. If the repo uses arc-skill-eval, add or update evals/evals.json with at least one positive path and one adjacent negative. Completion: the skill has runnable coverage or an explicit reason coverage is deferred.
6. Report the contract. Summarize created/changed files, invocation mode, disclosed references, eval coverage, and any assumptions the user should verify. Completion: the user can run or install the skill from the paths you provide.

Output contract

Produce file changes, not just advice, when the user asks to create or revise a skill. If information is missing, make the smallest safe assumption and call it out rather than blocking, unless the invocation mode or target skill name is ambiguous.