Skill Enhancer

Overview

Use this skill to turn a rough or existing skill into a sharper operational workflow. Prefer evidence from the target skill, nearby repo docs, local Claude/Codex skills, and known external skill patterns over generic advice.

When the user asks for a broad comparison, read references/source-patterns.md. Otherwise load it only if the current skill needs pattern selection or source-backed justification.

Workflow

1. Locate the target.

   • Find the skill directory, SKILL.md, metadata files, references, scripts, and any linked repo docs.
   • If porting from Claude, inspect frontmatter, allowed tools, command arguments, project rules, and related .claude/commands or .claude/agents files.
   • If no target exists, scaffold with the local skill-creator workflow and then edit.

2. Build the source map.

   • Target skill: what it claims, when it triggers, what workflow it enforces.
   • Local patterns: nearby successful skills, project-specific rules, templates, validation scripts.
   • External patterns: use Matt Pocock for grilling ambiguous plans, Superpowers for skill TDD and verification discipline, and Codex skill-creator for local structure and metadata.

3. Resolve ambiguity with a grill loop.

   • Ask one question at a time only when human judgment is required.
   • If the answer can be discovered from files, inspect files instead of asking.
   • Each question must include a recommended answer and the tradeoff it implies.
   • Continue until ambiguity is reasonably solved for the next edit, not until every theoretical edge case is exhausted.

4. Run the enhancement audit.

   • Trigger: frontmatter description says what the skill does and exactly when to use it.
   • Scope: states when to use it, when not to use it, and what inputs it expects.
   • Workflow: gives ordered actions, decision points, and completion criteria.
   • Disclosure: keeps SKILL.md lean; moves detailed patterns, examples, or domain rules to one-level references.
   • Tools: prefers deterministic scripts for repeated fragile operations.
   • Interaction: asks only useful questions; provides defaults or recommendations.
   • Safety: protects user edits, secrets, external side effects, and destructive actions.
   • Verification: names concrete checks before claiming completion.

5. Edit the skill.

   • Keep changes scoped to the requested skill unless related metadata or references must change.
   • Preserve useful source-specific practices while converting platform-specific assumptions.
   • Avoid duplicating large reference material in SKILL.md; link it instead.
   • Keep supporting files limited to files the agent should actually use.

6. Validate.

   • Run the Codex skill validator for Codex skills when available.
   • Search for unfinished scaffold markers, stale names, and broken relative references.
   • Verify agents/openai.yaml still matches the skill and mentions $skill-name in default_prompt.
   • For behavioral skills, pressure-test at least one realistic prompt. Use subagents only when available and useful.

7. Report completion.

   • Lead with changed files and validation evidence.
   • Summarize the source patterns used only as much as needed.
   • Do not claim the skill works until validation output or a concrete manual inspection supports it.

Output Shape

For small edits, return changed files plus validation. For larger enhancements, include:

• Source comparison: the 3-6 patterns that materially changed the skill.
• Patch summary: what changed in SKILL.md, references, scripts, or metadata.
• Verification: commands run and important results.
• Open risks: missing source access, untested behavior, or assumptions that still matter.