Skillify

Skillify

Codify a workflow only after it has worked. This is an explicit, approval-gated skill authoring flow, not autonomous transcript mining.

Use When

• A multi-step task finished green and the user says "skillify this", "save this as a skill", or similar.
• The same stable workflow has repeated enough times that a skill would reduce future mistakes.
• The user has approved creating or updating a specific skill.

Do Not Use When

• The workflow is still experimental or unresolved.
• The result depended on luck, private context, or unrepeatable local state.
• A helper script or README section would be simpler than a skill.
• A similar skill already exists; propose updating it instead.

Flow

1. Read the local skill-creator guidance before drafting.
2. Reconstruct the workflow spine:
   • user intent
   • required inputs
   • tool/file sequence
   • verification that proved it worked
   • failure modes discovered
3. Propose the slug, trigger description, scope, and whether the work updates an existing skill or creates a new one.
4. Show the full SKILL.md draft in chat and ask for explicit approval before writing.
5. Keep the skill concise:
   • frontmatter name and description
   • body under about 200 lines
   • long material in references/
   • deterministic repeated work in scripts/
   • reusable output files in assets/
6. Validate with:

python skills/skill-creator/scripts/quick_validate.py <skill-dir>

7. Report the created/changed paths and the validation result.

Frontmatter Template

---
name: <slug>
description: <when to use this skill, in one clear sentence>
---

Avoid nonportable frontmatter fields unless the target provider explicitly supports them.

Approval Rule

Each skill write needs its own explicit user approval immediately before the write. A broad "keep going" does not approve future skills, because skills auto-activate and can shape behavior across projects.

Failure Modes

• No stable workflow: stop and ask the user to describe the target workflow.
• Existing skill collision: propose an update instead of overwriting.
• Draft exceeds 200 lines: move details into references before asking approval.
• Validation fails: fix the skill package before reporting it as ready.