Skill Creator Meta

Skill — Skill Creator Meta (Iterative Skill Authoring & Optimization)

When this skill activates

When creating, authoring, testing, or optimizing MindForge skills. This is the meta-skill that governs the lifecycle of all other skills — from initial interview through eval-driven iteration to final registration. Use whenever a new skill needs to be created, an existing skill needs optimization, or skill quality needs systematic validation.

Core principle: Eval-driven iteration — skills are not "done" until they demonstrably outperform the no-skill baseline on realistic test cases.

Mandatory actions when this skill is active

Before skill creation begins

1. Interview for specification:

   • What does this skill do? (one-sentence purpose)
   • When should it trigger? (specific user intents, not vague categories)
   • What is the output format? (file, inline response, structured data)
   • What edge cases exist? (partial matches, ambiguous triggers, conflicting skills)
   • What dependencies does it have? (other skills via compose, external tools, MCP servers)

2. Validate against existing skills:

   • Check MANIFEST.md for trigger overlap (jaccard > 0.3 = conflict)
   • Verify the skill covers a genuinely distinct capability
   • If overlap exists: either extend existing skill or document why separation is needed

3. Define the skill schema constraints:

   • SKILL.md must be under 500 lines (enforced)
   • YAML frontmatter must include: name, version, min_mindforge_version, status, triggers
   • Optional: compose (list of skill dependencies activated alongside)
   • Body must include: "When this skill activates", "Mandatory actions when this skill is active" (Before/During/After), "Self-check before task completion"

During skill creation

Phase 1 — Draft:

• Write the SKILL.md following the schema exactly
• Triggers should be 8-15 phrases covering natural language variations
• Body should be action-oriented (imperative verbs, numbered steps)
• Include concrete examples where behavior might be ambiguous

Phase 2 — Create test cases:

.mindforge/evals/[skill-name]/
├── test-cases.json   # 2-3 realistic prompts complex enough to trigger loading
├── assertions.json   # objectively verifiable success criteria
└── benchmark.json    # aggregated results across iterations

Test case structure:

{
  "test_cases": [
    {
      "id": "tc-001",
      "prompt": "realistic user prompt that should trigger this skill",
      "context": "any relevant project state",
      "expected_behaviors": ["specific", "verifiable", "outcomes"],
      "should_trigger": true
    }
  ]
}

Phase 3 — Parallel evaluation:

• Run each test case WITH the skill loaded (experimental condition)
• Run each test case WITHOUT the skill loaded (baseline condition)
• Compare outputs on defined assertions
• Skill must demonstrably improve output quality over baseline

Phase 4 — Trigger optimization:

• Generate 20 trigger evaluation queries:
  • 10 that SHOULD trigger the skill (true positives)
  • 10 that should NOT trigger (true negatives — related but distinct intents)
• Test trigger matching accuracy
• Iterate on name/description/triggers up to 5 times until precision > 90%

Phase 5 — User review:

• Present: skill draft, test case outputs, benchmark scores, trigger accuracy
• Collect feedback on: missing behaviors, unnecessary behaviors, trigger gaps
• Document feedback for improvement phase

Phase 6 — Improvement:

• Generalize from specific feedback to patterns
• Bundle repeated helper logic into scripts/ directory within skill folder
• Rewrite sections that received negative feedback
• Ensure changes don't regress previously-passing test cases

Phase 7 — Final iteration:

• Rerun all test cases against improved skill
• Compare to previous iteration (must not regress)
• If regression: investigate, fix, rerun

After skill creation

1. Register in MANIFEST.md:

   • Add entry with name, version, triggers, status
   • Verify no trigger conflicts with existing entries

2. Verify file structure:

   .mindforge/skills/[skill-name]/
   ├── SKILL.md          # the skill definition (required)
   ├── scripts/          # helper scripts if needed (optional)
   └── README.md         # usage examples (optional)
   

3. Store eval artifacts:

   .mindforge/evals/[skill-name]/
   ├── test-cases.json
   ├── assertions.json
   ├── benchmark.json
   └── iterations/       # historical results per iteration
   

4. Final quality gate:

   • Skill passes all test cases
   • Trigger precision > 90% on 20-query eval set
   • Skill outperforms baseline on all assertions
   • No trigger conflicts in MANIFEST.md
   • Under 500 lines

Self-check before task completion

Before marking a skill creation task done:

• Did I interview for edge cases and dependencies?
• Did I create eval test cases (2-3 realistic prompts)?
• Did I run parallel comparison (with-skill vs baseline)?
• Did I optimize triggers with 20 evaluation queries (10 should, 10 should-not)?
• Did I iterate based on user feedback?
• Did I verify no trigger conflicts in MANIFEST.md?
• Is the skill under 500 lines?
• Does the final iteration outperform all previous iterations?