skill-creator

name	skill-creator
description	Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.

Skill Creator

Create, evaluate, improve, package, and maintain Claude skills. Use this skill when the user wants a new skill, an existing skill changed, skill performance measured, evals run, or a skill description optimized for triggering accuracy.

Core Loop

1. Determine where the user is: new skill, draft review, eval run, improvement, description optimization, packaging, or troubleshooting.
2. Capture intent and confirm the required input contract before writing.
3. Draft or edit the skill folder.
4. Create realistic test prompts when testing is useful.
5. Run skill-vs-baseline evals when tooling supports it; otherwise run a smaller qualitative review.
6. Grade outputs, gather user feedback, and improve the skill.
7. Repeat until the user is satisfied, objective checks pass, or progress stalls.
8. Package or present the final skill when requested and supported.

Required Input Contract

Before writing a new skill, identify:

• what the skill enables Claude to do
• when it should trigger, including user phrases and contexts
• expected output format or artifacts
• dependencies, tools, scripts, reference docs, or assets
• safety boundaries and prohibited behavior
• whether evals are useful for the task type

If the conversation already contains the workflow to turn into a skill, extract the sequence, tools, corrections, input/output formats, and success criteria from the conversation before asking follow-up questions.

Skill Folder Contract

Use this structure unless the existing skill already has a compatible layout:

skill-name/
  SKILL.md
  scripts/
  references/
  assets/
  evals/evals.json

• SKILL.md is required and holds frontmatter plus concise operating guidance.
• scripts/ holds deterministic or repetitive executable helpers.
• references/ holds lazy-loaded detailed docs.
• assets/ holds templates, icons, examples, or output resources.
• evals/evals.json holds skill test prompts when evals are used.

Keep SKILL.md concise. Move bulky examples, schemas, deep explanations, templates, troubleshooting detail, and long playbooks into references/.

Frontmatter Requirements

Every SKILL.md must include YAML frontmatter with:

• name: stable skill identifier, usually matching the folder name.
• description: the primary trigger mechanism. Include what the skill does and when to use it.

Descriptions should be specific and slightly trigger-forward. Include likely user intents and contexts, not only a generic capability label. Do not hide triggering rules only in the body.

Safety and Security Constraints

• Do not create misleading skills.
• Do not add malware, exploit code, credential harvesting, data exfiltration, or unauthorized-access workflows.
• Do not include secrets, credentials, private context, local user paths, session records, memory exports, personal identifiers, or real tokens in skill files.
• Do not overfit instructions to one example when the skill should generalize.
• Preserve existing skill name and frontmatter identity when improving an installed skill unless the user explicitly asks for a rename.
• Treat bundled scripts as code: keep them scoped, inspectable, and testable.

Creation Workflow

1. Capture intent: confirm capability, trigger contexts, output format, dependencies, success criteria, and safety limits.
2. Research: inspect existing files, examples, tools, references, and related skills. Ask only the missing questions that affect the skill design.
3. Draft structure: create SKILL.md; add scripts/, references/, assets/, and evals/ only when they carry real value.
4. Write instructions: use imperative, general, reusable guidance. Explain why important rules matter.
5. Add minimal examples: include only examples needed for correct invocation or output shape; move longer examples to references.
6. Validate: check frontmatter, links, file layout, safety constraints, and eval schema when present.
7. Evaluate: run test prompts when meaningful; compare against baseline for measurable skills.
8. Iterate: improve based on transcript evidence, grader output, benchmarks, and user feedback.

Full phase detail: references/workflow-deep-dive.md.

Validation Rules

• Do not claim eval success without evidence.
• Do not use /skill-test or another testing skill for the skill-creator eval loop.
• For eval runs, create a sibling workspace named <skill-name>-workspace/ and organize by iteration-N/eval-N/.
• For new skills, baseline is without_skill.
• For improving existing skills, snapshot or otherwise preserve the old version before comparing.
• Capture timing from completion notifications immediately when available.
• Use eval-viewer/generate_review.py for human review when browser/static output is available; do not invent custom review HTML.
• Use agents/grader.md, agents/analyzer.md, and agents/comparator.md only for their documented roles.

Detailed eval, grading, viewer, failure-mode, and security guidance: references/validation-and-safety.md, references/eval-workflow.md, and references/schemas.md.

Minimal Eval Example

{
  "skill_name": "example-skill",
  "evals": [
    {
      "id": 1,
      "prompt": "User's task prompt",
      "expected_output": "Description of expected result",
      "files": []
    }
  ]
}

Full schemas: references/schemas.md.

Output Expectations

When creating or modifying a skill, report:

• skill folder and files created or changed
• trigger behavior and output contract
• safety constraints included
• validation performed and results
• evals run, if any, with evidence
• known limitations and recommended next iteration

If packaging is requested and supported, use scripts/package_skill.py and point to the created package artifact.

Environment Variants

• Claude Code/Codex with subagents: run parallel skill-vs-baseline evals when useful.
• Claude.ai without subagents: run qualitative tests one at a time and ask for user review inline.
• Headless/Cowork-like environments: generate a static review artifact instead of a browser server.

Variant details and packaging notes: references/workflow-deep-dive.md.

Reference Map

• references/workflow-deep-dive.md - full creation workflow, eval phase details, environment variants, packaging, and advanced decisions.
• references/frontmatter-and-structure.md - frontmatter examples, folder layout, naming, output contracts, and resource placement.
• references/authoring-patterns.md - writing patterns, long templates, description guidance, progressive disclosure, and reusable instruction forms.
• references/validation-and-safety.md - extended validation checks, failure modes, eval mechanics, security caveats, and troubleshooting.
• references/examples.md - end-to-end skill creation and improvement examples.
• references/eval-workflow.md - existing detailed eval-and-review procedure.
• references/description-optimization.md - existing trigger-description optimization procedure.
• references/schemas.md - JSON schemas for eval, grading, timing, and metrics.
• references/skill-template-reference.md - existing anatomy diagram and progressive-disclosure reference.

Specialized agents:

• agents/grader.md - evaluate assertions against outputs.
• agents/comparator.md - blind A/B comparison.
• agents/analyzer.md - analyze why one version beat another.