Skillify

Capture a process that worked well and convert it into a reusable skill file. Use this to scale team workflows without rewriting instructions every time.

Use when

• The same workflow appears repeatedly across sessions.
• A process has stable inputs, ordered steps, and clear success criteria.
• You want to convert tacit know-how into a shared SKILL.md.

Do not use when

• The process is one-off, exploratory, or still unstable.
• Key steps are unknown or depend on hidden tribal context.
• The user only wants a quick summary, not a reusable artifact.

Input contract

• workflow_name (required): short name for the new skill.
• source_scope (optional): current_task (default) or recent_session.
• target_location (optional): default skills/<slug>/SKILL.md.
• required_args (optional): known runtime arguments to expose.
• confirm_before_write (optional): default true.

Parsing rules (priority order)

1. Structured args override everything.
2. Else infer from user phrasing:
   • "make a skill for X" -> workflow_name=X
   • "from this session" -> source_scope=recent_session
   • "save under ..." -> target_location=...
3. Defaults:
   • source_scope=current_task
   • target_location=skills/<slug>/SKILL.md
   • confirm_before_write=true
4. If workflow_name is missing, ask for a concrete name.

Canonical workflow

1. Extract process skeleton
   • Goal, inputs, step sequence, outputs, verification criteria.
2. Identify constraints
   • Required tools, risk boundaries, irreversible actions, approvals.
3. Draft skill frontmatter + body
   • Include Use when, Do not use when, input contract, workflow, guardrails, and output format.
4. Review with user
   • Show full draft and request approval before writing when confirm_before_write=true.
5. Write skill artifact
   • Create directory/file and persist final SKILL.md.
6. Post-write validation
   • Re-read file and confirm it is syntactically intact and complete.

Guardrails

• Never overwrite an existing skill silently.
• Avoid over-broad automation promises; keep the skill bounded.
• Prefer minimal required tool surface over wildcard permissions.
• Mark ambiguous steps explicitly instead of guessing.
• Keep language implementation-oriented, not aspirational.

Output format

Always return:

• status: drafted | written | needs_user_input
• workflow_name: normalized name
• target_location: final path
• draft_summary: what the generated skill covers
• open_questions: unresolved ambiguities
• next_action: required unless status=written

Example

• workflow_name="release-pr-handoff" source_scope="recent_session" target_location="skills/release-pr-handoff/SKILL.md"