Skill Creation

Skill Creation (planning + creation)

This skill runs two flows: planning (/skill-plan) and creation (/skill-create). Review is handled separately by the skill-reviewing skill. Keep the main conversation clean; heavy work runs in agents via Agent Runner with registry-based resolution.

Planning Flow (wizard)

1. Confirm goal: new skill package vs update/version bump (capture desired version target if known).

2. Capture responsibilities + success criteria; list primary use cases.

3. Command UX: required args, options, --help text; defaults for --name, behavior of --from (seed from existing skill/plan), and --template (plan skeleton file). Enforce naming convention: file path .claude/commands/<name>.md, frontmatter name: <name> (no slash), user invocation /name (slash added by system), default version 0.1.0.

4. Agents: propose candidate agents, inputs/outputs, fenced JSON envelopes, and registry versions.

5. File layout: commands under .claude/commands, skills under .claude/skills/<name>/SKILL.md, agents under .claude/agents, references/templates under .claude/references (if needed) for easy testing before packaging. Default versions for commands/skills/agents: 0.1.0 unless explicitly overridden.

6. Keep agent set minimal by default; only add extra agents when scope justifies it. Prefer clear roles (e.g., intake, fix, PR) over many micro-agents.

7. Statusing: mark plan as Preliminary, then Proposed after user confirms.

8. Review Gate (Pre-Approval): Before marking a plan as Approved, invoke all three review agents to validate the proposed architecture against v0.5 guidelines:

   • skill-architecture-review — Validates two-tier pattern, response contracts, agent design
   • skill-implementation-review — Validates fenced JSON, hook patterns, security considerations
   • skill-metadata-storage-review — Validates frontmatter requirements, storage conventions

   All three must return success: true with zero errors before the plan can be marked Approved. Warnings are advisory and do not block approval. If any review returns errors, address the issues in the plan before re-running the review gate.

9. Generate 40–80 line overview with full plan path when user approves (Approved). Planning must not create artifacts—only write/update a plan file. Creation happens only via /skill-create.

Implementation notes

• Use Agent Runner; resolve agents by name+version from .claude/agents/registry.yaml.
• Store plans by default in plans/<name>.md; allow override when user supplies a path or --template. Do not overwrite request/input files—write to a new plan file/path when invoked with a request doc.
• If --from is set, load referenced artifact(s) and summarize key facts into the plan.
• Maintain concise outputs; omit tool traces.

Agent Runner invocation (v0.5 contract)

agent: "skill-planning-agent"
params:
  goal: "new|update"
  target_name: "<skill name>"
  plan_path: "<plans/path.md>"
  from_paths: ["<paths from --from if any>"]
  template_path: "<template if provided>"
version_constraint: "0.x"
timeout_s: 120

Agent output (fenced JSON, minimal envelope)

{
  "success": true,
  "data": {
    "summary": "Concise overview",
    "plan": { "path": "<dest>", "status": "Preliminary|Proposed|Approved" },
    "actions": ["next step", "next step"],
    "open_questions": ["q1", "q2"]
  },
  "error": null
}

If plan incomplete, include open questions and next actions; otherwise provide summary + full plan path.

Review Gate invocation (parallel execution)

Run all three review agents in parallel when transitioning from Proposed → Approved:

# Architecture review
agent: "skill-architecture-review"
params:
  target_type: "plan"
  target_path: "{{plan_path}}"
  check_two_tier: true
  check_contracts: true
  check_naming: true
version_constraint: "0.x"
timeout_s: 60

# Implementation review
agent: "skill-implementation-review"
params:
  target_path: "{{plan_path}}"
  check_hooks: true
  check_dependencies: true
  check_security: true
version_constraint: "0.x"
timeout_s: 60

# Metadata & storage review
agent: "skill-metadata-storage-review"
params:
  target_type: "plan"
  target_path: "{{plan_path}}"
  check_registry: false  # Plan doesn't have registry entries yet
version_constraint: "0.x"
timeout_s: 60

Review Gate output aggregation

Collect results from all three agents and emit consolidated status:

{
  "success": true,
  "data": {
    "review_gate_passed": true,
    "reviews": {
      "architecture": { "passed": true, "errors": 0, "warnings": 1 },
      "implementation": { "passed": true, "errors": 0, "warnings": 0 },
      "metadata_storage": { "passed": true, "errors": 0, "warnings": 2 }
    },
    "blocking_issues": [],
    "advisory_warnings": ["...", "..."],
    "plan_status": "Approved"
  },
  "error": null
}

If review_gate_passed: false, the plan remains at Proposed status until issues are resolved.

Creation Flow

1. Input: approved plan path.
2. Generate stubs for command/skill/agents/references at version 0.1.0 using naming convention:
   • File path: .claude/commands/<name>.md
   • Frontmatter name: <name> (no slash)
   • User invocation: /name (slash added by system)
3. Populate references listed in the plan (e.g., required reference files).
4. Update registry with new agents and skill dependency constraints (including manage-worktree when git ops are needed).

Storage

Safety

• Creation of plan files is allowed; avoid destructive repo changes.
• Keep outputs concise; no tool traces in main conversation.
• Base branch/worktree defaults: honor .claude/config.yaml if present (base_branch, worktree_root); otherwise default to main and a repo-named worktree root. For git operations, instruct reuse of manage-worktree skill.
• Registry expectations: include manage-worktree dependency when git ops are needed; use constraints like manage-worktree: "0.x". Ensure new agents/skills are added with version 0.1.0 by default.

Reference Documents

Plans and generated artifacts MUST comply with these normative documents:

Review Agents

The review gate uses these specialized agents (from skill-reviewing skill):