skill-jit

name

skill-jit

description

Researches any capability and crystallizes it into a reusable skill with the right structural pattern (tool wrapper, generator, reviewer, pipeline, or inversion). Trigger when the user asks to "create a skill", "make a skill for", "turn this into a skill", or "learn how to use X". Also trigger proactively when: (1) the task involves unfamiliar CLI tools, APIs, libraries, or workflows that need research; (2) the task would benefit from a structured, repeatable workflow — such as code review with checklists, consistent document generation, multi-step deployment with approval gates, or project scaffolding with configuration interviews; (3) a previously generated skill fails, produces poor results, or a referenced skill is missing. Do NOT trigger for tasks you can already handle reliably without a structured workflow.

Create Skill

Research, verify, and crystallize any capability into a reusable skill. Also fix existing skills that have failed.

Process

1. Parse Input

Extract the task description from $ARGUMENTS. If empty, ask what the user wants the skill to do.

Flags:

• --dry-run: Preview what would be created without writing files

2. Detect Mode

Fix mode — if the input mentions:

• An existing skill name AND a failure/error/problem
• Keywords like "failed", "broken", "error", "doesn't work", "fix", "update"

Create mode — otherwise (default).

3. Capture Intent

Quickly establish what the skill needs to do. Check if the current conversation already contains a workflow to capture — if so, extract answers from context first.

Key questions (ask only what's not already clear):

1. What should this skill enable you to do?
2. When should it trigger? (what user phrases/contexts)
3. What's the expected output?
4. Are there edge cases or constraints to consider?

Pattern hint (infer from context, don't ask unless ambiguous):

• Wraps a tool/API → Tool Wrapper
• Produces consistent artifacts → Generator
• Checks/audits/validates → Reviewer
• Needs user decisions before acting → Inversion
• Strict sequential steps with irreversible actions → Pipeline

If the pattern is obvious from the description, note it. If ambiguous, briefly ask.

Proceed directly if the description or conversation context already answers these. Don't over-interview.

4. Determine Output Directory

For Claude Code, skills are saved to .claude/skills/ relative to the current working directory (the project root). If you are not working in a Claude agent, please determine an appropriate directory to save the generated skill files, and ensure the writer agent is informed of this location in step 5.

5. Execute

Create Mode

1. Generate: Spawn the skill-writer agent. The writer handles everything — it decomposes the task, spawns researchers as needed, and generates skill files. Pass the resolved output directory so it knows where to write.

   Task(
     subagent_type: "skill-jit:skill-writer",
     prompt: "Create skill(s) for: [original description]\n\nOutput directory: [resolved output directory from step 4]\nPattern hint: [detected pattern from step 3, or 'auto' if not determined]",
     description: "Create skill"
   )
   

2. Report: Output the writer's completion report verbatim. This is critical because skill-jit runs in a forked context — the main context only sees this output. The writer's report includes skill paths, missing references, the original task, and instructions to load the skill.

Fix Mode

1. Locate skill: Search ./.claude/skills/ or other skill directories for the mentioned skill

2. Read current skill: Load the existing SKILL.md

3. Fix: Spawn skill-writer in fix mode with current skill content + error context:

   Task(
     subagent_type: "skill-jit:skill-writer",
     prompt: "Fix this skill. Error: [error]. Current skill: [content]. User request: [description]\n\nOutput directory: [same directory as the existing skill]",
     description: "Fix skill"
   )
   

4. Report: Display what was changed

6. Dry-Run Mode

When --dry-run is provided:

1. Complete research phase normally
2. Have writer generate content but NOT write files
3. Display preview of what would be created
4. Allow user to review before actual creation

Tips

• Be specific about what you want — the more detail, the better the skill
• Mention specific tools if you have a preference
• For failures, include the actual error message
• Use --dry-run to preview before committing to file creation