FORGE-skill-creator

skill-creator (O_Ψ Verification Layer)

Purpose

Bootstrap, design, and package new skills.

Use When

1. Capturing user intent for a brand new skill from conversation history.
2. Conducting interviews to identify edge cases, input/output requirements, and dependencies for a new skill.
3. Drafting the initial structure, frontmatter, and instructions for a new SKILL.md file.
4. Packaging finished skill folders into .skill distributables.

Do Not Use When

1. Auditing the active skill portfolio for collisions or stale documentation (use arifos-recursive-audit instead).
2. Running quantitative benchmarks or evaluating prompt traces (use 080-eval-benchmark-test instead).
3. Linting trigger names or checking vague verb usage (use skill-trigger-linter instead).

Inputs

• User Concept: The high-level intent or workflow specified by the user.
• System Target: The proposed skill workspace directory.

Procedure

1. Capture Intent: Analyze the user request and extract target tools, sequence steps, and expected outputs.
2. Define Trigger Boundaries: Design preliminary "Use when" and "Do not use when" rules.
3. Draft SKILL.md: Compose the frontmatter, purpose, triggers, inputs, procedure, postconditions, failure modes, and telemetry structure.
4. Verification Setup: Propose 2-3 realistic test prompts for verification.
5. Floor-Tier SEAL Gate (MANDATORY — see §Floor-Tier SEAL Gate below): Classify the drafted skill by primary floor touch. Route to the correct SEAL authority before proceeding.
6. Package Skill: Run scripts/package_skill.py to compile the folder structure into a distributable package.

Floor-Tier SEAL Gate (MANDATORY)

Before any newly drafted skill is sealed into the live meta-mesa-skill-atlas, the creator MUST classify the primary floor touch and route to the correct authority:

Floor classification rule: Use the floor_scope declared in the skill's own frontmatter. If no frontmatter, default to F2. If the skill touches multiple floors, use the highest-numbered floor as the governing tier.

Grader output requirement: The Grader must emit pass_rate ≥ 0.95 for the skill to qualify for any SEAL path. Skills with pass_rate < 0.95 → return to Step 3 (redraft).

Verification of the Grader: The Grader agent evaluates the skill against realistic test prompts and outputs grading.json. The Grader does NOT have SEAL authority. It is a quality gate, not a verdict authority. Verdict authority stays with the floor tier above.

Telemetry update: On reaching the Floor-Tier SEAL Gate, update the running telemetry:

{
  "floor_tier_reached": "<F1|F2|F5>",
  "sealing_authority": "<888|human|autonomous>",
  "grader_pass_rate": <0.0-1.0>,
  "hold_code": "<if applicable>"
}

Postconditions

1. A valid skill structure (including SKILL.md) is successfully staged in the target folder.
2. The skill complies with the baseline structural layout standards.
3. The package is compiled without system compile errors.
4. The Floor-Tier SEAL Gate was observed and the correct authority was used.

Failure Modes & Escalation

• Intent Ambiguity: The user request is too vague to extract structured boundaries. Action: Pause generation and present clarifying questions regarding concrete tools and expected formats.

Telemetry per Run

{
  "skill_name": "skill-creator",
  "version": "1.1.0",
  "trigger_phrase": "{{trigger_phrase}}",
  "selected_reason": "{{selected_reason}}",
  "latency_ms": 0,
  "token_in": 0,
  "token_out": 0,
  "commands_run": 0,
  "artifacts_written": 1,
  "postcondition_pass": false,
  "floor_tier_reached": "{{F1|F2|F5}}",
  "sealing_authority": "{{888|human|autonomous}}",
  "grader_pass_rate": 0.0,
  "human_approval_required": false,
  "hold_code": "{{hold_code}}"
}

Recursive Scorecard

• Activation Precision: [0.0 - 1.0] (Target: >0.95)
• Task Completion Rate: [0.0 - 1.0] (Target: >0.95)
• Rollback Safety: [0.0 - 1.0] (Target: 1.00)
• Context Efficiency: [0.0 - 1.0] (Target: >0.90)
• Doc Freshness: [0.0 - 1.0] (Target: 1.00)
• Cross-Skill Collision Rate: [0.0 - 1.0] (Target: 0.00)
• Human Trust Score: [0.0 - 1.0] (Target: >0.98)