skill-builder

name	Skill Builder
description	Create new Anthropic Agent Skills, test them with eval-driven iteration, and optimize descriptions for triggering accuracy. Use when user asks to create a new skill, improve an existing skill, test skill quality, optimize skill descriptions, add skill functionality, or wants to build a custom workflow. Also use when analyzing why a skill under-triggers or produces poor results.

Skill Builder

Create, test, and iteratively improve Anthropic Agent Skills.

Quick Start

# Create a new skill
scripts/create_skill.sh --project my-skill "Does X. Use when user asks for Y."

# Validate a skill
scripts/validate_skill.sh my-skill

# Test skill triggers
scripts/test_triggers.sh /path/to/skill trigger_eval_set.json

# Run full eval (with-skill vs baseline)
scripts/run_skill_eval.sh /path/to/skill evals.json

# Optimize description for better triggering
scripts/optimize_description.sh /path/to/skill

When to Use This Skill

• User asks to "create a skill" or "make a new skill"
• User wants to test or evaluate a skill's quality
• User wants to optimize a skill's description for better triggering
• User says "the skill isn't triggering" or "the skill gives bad results"
• During skill execution: When you discover issues in ANY skill
• Self-improvement: When a script fails or could be enhanced

Core Workflow: Draft → Test → Improve → Repeat

The key to good skills is iteration. Don't just write and ship — test, get feedback, improve.

1. Capture Intent

Understand what the skill should do:

• What should this skill enable Claude to do?
• When should this skill trigger? (what user phrases/contexts)
• What's the expected output format?
• Are the outputs objectively verifiable? (→ use automated evals)

2. Create Structure

scripts/create_skill.sh --global|--project <name> "<description with triggers>"

Creates: SKILL.md, scripts/example_script.sh, resources/reference.md

3. Write the Skill

Description (most critical): The description determines triggering. Be slightly "pushy" — list scenarios where the skill should activate, including edge cases. Claude tends to under-trigger, so err on the side of being explicit.

Instead of: "Manages database migrations."

Write: "Manage database migrations including schema changes, rollbacks, and seed data. Use when user mentions migrations, database schema, ALTER TABLE, asks to update the database, or wants to roll back a deployment, even if they don't explicitly say 'migration'."

Instructions: Explain the why behind what you're asking the model to do. Today's models are smart — they generalize better from reasoning than from rigid rules. If you find yourself writing ALWAYS or NEVER in caps, reframe it as an explanation of why it matters.

Progressive disclosure: Keep SKILL.md under 300 lines. Move detailed references to resources/, automation to scripts/.

4. Write Test Cases

Create 2-3 realistic test prompts in evals.json:

{
  "skill_name": "my-skill",
  "evals": [
    {
      "id": 1,
      "prompt": "realistic user prompt with context and details",
      "expected_output": "what good output looks like",
      "assertions": ["Output contains valid JSON", "No placeholder text remains"]
    }
  ]
}

See resources/schemas.md for full schema.

5. Run Evals

scripts/run_skill_eval.sh /path/to/skill evals.json

This spawns parallel with-skill and baseline runs, captures timing/tokens, saves to workspace/iteration-N/.

6. Grade Results

Use the grader agent (agents/grader.md) to evaluate assertions. For programmatic assertions, write and run a script rather than eyeballing — it's faster and reusable.

7. Improve and Repeat

Read transcripts (not just final outputs). Look for:

• Wasted effort the skill is causing (remove those instructions)
• Patterns across test cases (bundle repeated scripts)
• Feedback from users (generalize, don't overfit to examples)

Rerun with --iteration N+1. Keep going until feedback is empty or you're not making progress.

8. Optimize Description

After the skill is stable:

scripts/optimize_description.sh /path/to/skill

Generates 20 trigger eval queries, splits 60/40 train/test, iteratively improves the description, and selects by test accuracy to avoid overfitting.

9. Validate

scripts/validate_skill.sh my-skill

Available Scripts

Script	Purpose
create_skill.sh	Scaffold new skill structure
validate_skill.sh	Check skill follows specification
test_triggers.sh	Test if skill triggers for given prompts
run_skill_eval.sh	Run with-skill vs baseline comparison
optimize_description.sh	Iteratively improve description triggering

Agent Prompts

Use these when spawning subagents for evaluation:

Agent	When to Use
agents/grader.md	Evaluate assertions against outputs with evidence
agents/comparator.md	Blind A/B comparison between two outputs
agents/analyzer.md	Find patterns hidden by aggregate stats

Resources

Resource	Content
schemas.md	JSON schemas for evals, grading, triggers
skill_template.md	Templates and naming conventions
self_improvement_guide.md	How skills self-improve
long_running_patterns.md	Multi-session continuity

Best Practices

1. Pushy descriptions — Combat under-triggering by listing specific scenarios, including "even if they don't explicitly ask for X"
2. Explain the why — Reasoning > rigid rules. Models generalize from understanding, not from ALL CAPS INSTRUCTIONS
3. Generalize from feedback — Don't overfit to test cases. Your skill will be used across many different prompts
4. Keep it lean — Remove instructions that aren't pulling their weight. Read transcripts to find unproductive steps
5. Bundle repeated work — If every test run independently writes a similar helper script, bundle it in scripts/
6. Progressive disclosure — SKILL.md ≤300 lines. Details in resources/, automation in scripts/
7. Test before shipping — Run at least one eval iteration. Untested skills accumulate subtle issues

Notes

• Global skills (~/.claude/skills/) are available across all projects
• Project skills (.claude/skills/) are specific to one project
• Scripts run without loading into context (saves tokens)
• Resources load only when needed