skill-creator-agent

L1 Improvement

• Reorganized the skill into the Skill Forge section pattern and added joint outputs for skill specs and agents.
• Applied prompt-architect ceilings, constraint extraction, and adversarial validation to both artifacts.

---

LIBRARY-FIRST PROTOCOL (MANDATORY)

Before writing ANY code, you MUST check:

Step 1: Library Catalog

• Location: .claude/library/catalog.json
• If match >70%: REUSE or ADAPT

Step 2: Patterns Guide

• Location: .claude/docs/inventories/LIBRARY-PATTERNS-GUIDE.md
• If pattern exists: FOLLOW documented approach

Step 3: Existing Projects

• Location: D:\Projects\*
• If found: EXTRACT and adapt

Decision Matrix

Match	Action
Library >90%	REUSE directly
Library 70-90%	ADAPT minimally
Pattern exists	FOLLOW pattern
In project	EXTRACT
No match	BUILD (add to library after)

---

STANDARD OPERATING PROCEDURE

Purpose

Create skills that come packaged with a dedicated specialist agent, including SKILL.md content, agent prompt, and integration hooks for registry use.

Trigger Conditions

• Positive: request to build a reusable skill that relies on a specialist agent, or to retrofit an existing skill with a dedicated agent.
• Negative/reroute: generic prompt tuning (prompt-architect), pure skill scaffolding (skill-builder), or standalone agent creation (agent-creator).

Guardrails

• Deliver both artifacts: SKILL.md and agent spec with frontmatter; no status-only responses.
• Use English outputs with explicit confidence ceilings.
• Ensure agent scope matches skill scope; avoid duplicated capabilities.
• Run at least one validation scenario that exercises the skill+agent pair.

Execution Phases

1. Scoping: Capture skill purpose, triggers, constraints, and domain; classify HARD/SOFT/INFERRED.
2. Skill Drafting: Author SKILL.md with SOP, guardrails, integrations, and IO contracts following Skill Forge.
3. Agent Drafting: Create agent prompt, tools, and refusal policy aligned to the skill.
4. Validation: Test the pair together (happy + edge cases) and record evidence with ceilings.
5. Delivery: Provide artifacts, routing metadata, and next-step improvements.

Pattern Recognition

• Tool-driven skills → agent should manage tool orchestration and error handling.
• Review/analysis skills → agent must enforce evidence and ceiling discipline.
• Integration-heavy skills → include latency and timeout guidance in both artifacts.

Advanced Techniques

• Use few-shot examples shared between skill and agent for consistency.
• Add capability tags so agent-selector can route correctly.
• Capture reusable snippets for meta-tools to compose in other skills.

Common Anti-Patterns

• Skill and agent scopes diverge, causing confusion.
• Missing tests for the combined flow.
• Absent confidence ceilings or unclear output formats.

Practical Guidelines

• Keep the agent persona pragmatic; focus on behaviors and contracts.
• Explicitly document inputs/outputs and refusal criteria in both artifacts.
• Provide registry path suggestions and tagging.

Cross-Skill Coordination

• Upstream: prompt-architect for clarity; skill-builder/skill-forge for structure.
• Downstream: agent-selector for routing; recursive-improvement for ongoing tuning; meta-tools for tool wiring.

MCP Requirements

• Document any MCP/tool dependencies and tags; tag WHO=skill-creator-agent-{session}, WHY=skill-execution if memory used.

Input/Output Contracts

inputs:
  skill_goal: string  # required objective
  domain: string  # required domain or category
  tools: list[string]  # optional tools/MCP servers
  constraints: list[string]  # optional constraints
outputs:
  skill_spec: file  # SKILL.md content
  agent_spec: file  # agent prompt with frontmatter
  validation: file  # tests and results for the pair

Recursive Improvement

• Use recursive-improvement on both artifacts together; stop when deltas < 2% or risks logged.

Examples

• Build a log-analysis skill with an observability agent handling query composition and summarization.
• Pair a security-advisory skill with an agent that enforces CVE evidence and refusal policy.

Troubleshooting

• Agent oversteps scope → tighten persona and refusal rules.
• Skill unclear → rerun constraint extraction and simplify outputs.
• Tool errors → add retries and error handling in agent prompt; document limits in skill.

Completion Verification

• SKILL.md and agent spec delivered with aligned scope.
• Validation evidence recorded with ceilings.
• Tool/MCP dependencies documented; routing metadata provided.
• Examples provided demonstrating combined use.

Confidence: 0.70 (ceiling: inference 0.70) - Skill Creator Agent SOP rewritten with Skill Forge cadence and prompt-architect ceilings.