Create Agent Skill

Boundaries

• Answer planning-only requests — plans, designs, critiques, outlines, requirements, skill-shape discussions — in conversation; for mixed requests, resolve the plan first and write files once the implementation path is clear.
• Treat named tools, skills, frameworks, and workflows as context for the plan; switch to file creation only on the user's explicit ask.

Workflow

1. Confirm scope.
   • Identify the concrete artifact to create or update: skill directory, SKILL.md, references, scripts, assets, or inventory docs.
   • If file-writing intent is ambiguous and edits would be surprising, ask for confirmation before editing.
   • Identify the task, workflow, project convention, or domain the skill covers.
   • When the skill is a shortcut — instructions the user retypes each time and wants behind a name — write it directly and stop: name per step 2, user-invoked frontmatter with a one-line description, and the instructions themselves as the body.
   • Capture the prompts, contexts, files, and workflows that should trigger it.
   • Note adjacent requests and near-misses that should not trigger it.
   • Decide whether the skill needs instructions only or also references, scripts, assets, templates, or sample files.
   • Ground the skill in real expertise — completed tasks, user corrections, project docs, runbooks, schemas, review comments, issues, or patches — and collect the reference material it should preserve.
   • Before encoding a lesson as instructions, check whether a mechanism can enforce it instead — a lint rule, a type, a runtime check, a script. Encode it in the skill's own artifacts (a bundled script, a hook) when it fits there and skip the sentence; when the mechanism belongs in the target repo, recommend it rather than editing beyond the skill. Prose is the fallback, not the default.