| name | skill-forge |
| description | Crystallize a captured workflow journal into a validated Claude skill. Use this whenever the user says "turn this into a skill", "make this repeatable", "build a skill from this", or when a workflow-capture worklog is ready to become a skill. Also use to improve or extend an existing produced skill from a new worklog. |
Skill Forge
Turn a WORKLOG.md into a skill. The journal's tags map directly onto skill sections, so synthesis is mostly routing.
Step 1: Read the worklog
Read the full _worklogs/<date-slug>/WORKLOG.md. If the user references a workflow without a path, find the matching worklog.
Step 2: Route by tag
- DECISION + SEQUENCE -> the procedure: ordered, imperative steps.
- CORRECTION + DEADEND -> guardrails: "do not", "always", warnings, each with the one-line reason.
- CONVENTION + DISCOVERY -> reference files and compatibility notes.
A DEADEND that blocked progress usually becomes an early SEQUENCE step in the procedure so it cannot recur.
Step 3: Draft SKILL.md
Read references/skill-anatomy.md first. Then produce:
- name: kebab-case identifier.
- description: the trigger. Make it pushy and list concrete contexts; Claude tends to undertrigger skills. Put all "when to use" info here, not in the body.
- body: imperative voice, under ~500 lines. Explain why a rule matters rather than stacking bare MUSTs.
- references/: push variant-specific or long details out of the body. Files over ~300 lines get a table of contents.
- scripts/ and assets/ only if the workflow has deterministic steps or output templates.
Step 4: Validate
Adapt to the environment.
- Subagents available (e.g. Claude Code): run 2 to 3 realistic test prompts, baseline vs with-skill, grade, review the results.
- No subagents (e.g. a chat client): run test prompts inline, one at a time, using the skill yourself. Skip quantitative benchmarking; rely on qualitative review.
Iterate: run -> review -> revise SKILL.md -> rerun. Optimize the description for triggering once stable.
Step 5: Archive and link
Move the source worklog to skills/<name>/.source/WORKLOG.md for provenance. Archive, never delete. Symlink or copy the produced skill into the runtime skill path so it loads.
Step 6: Package (if available)
If a packaging script is available, package the skill and hand back the artifact path.
