Skill Crafting

Skill Crafting

Create effective, discoverable skills that work under pressure.

When to Use

Creating:

• "Create a skill for X"
• "Build a skill to handle Y"

From Session History:

• "Create a skill from this session"
• "Turn what we just did into a skill"
• "Can the database setup we did become a skill?"
• "Could we create a skill from this?" (evaluate first)
• "Should this be a skill?" (evaluate first)

Fixing:

• "This skill isn't working"
• "Why isn't this skill triggering?"
• "Skill didn't trigger when it should have"

Analyzing:

• "Analyze my skill for issues"
• "Run skill analysis"
• "Check this skill's quality"
• "Audit all my skills"
• "Check character budget across skills"

Analyzing a Skill

When user asks to analyze a skill:

1. Run scripts first when available for mechanical checks:

   python3 ${CLAUDE_SKILL_DIR}/scripts/analyze-all.py path/to/skill/
   

   If scripts fail because of environment or tooling issues, state the blocker clearly and continue with manual review.

2. Read the skill files for qualitative review:

   • Read SKILL.md
   • Read REFERENCES.md (if exists)
   • Read directly linked references/, scripts/, or agents/openai.yaml files when they materially affect behavior

3. Provide holistic feedback covering:

   • Script results (CSO, structure, tokens) or explicit blocker if scripts could not run
   • Does allowed-tools match what the skill needs to do?
   • Does the analysis cover the files that actually define the skill's behavior?
   • Is the workflow clear and actionable?
   • Are references appropriate and sized correctly?
   • Missing sections or anti-patterns?

4. Give verdict with prioritized recommendations

Validation Scripts

Quick start:

python3 ${CLAUDE_SKILL_DIR}/scripts/analyze-all.py ~/.claude/skills/my-skill/
python3 ${CLAUDE_SKILL_DIR}/scripts/check-char-budget.py ~/.claude/skills/

Creating from Current Session

When user asks to create a skill from the current session:

1. Reflect on the conversation — you already have full context
2. Assess skill-worthiness using criteria below
3. If worthy: Generate SKILL.md using methodology in this skill
4. If not: Explain why (one-off, too scattered, etc.)

Skill-Worthiness Criteria

Quick Assessment

Before creating, answer:

1. What pattern repeats? (e.g., "set up auth with tests")
2. What would break without the skill? (steps someone might skip)
3. Who else would use this? (just me? team? public?)

If you can't answer these clearly → probably not skill-worthy.

For Evaluative Questions

When user asks "could this be a skill?" or "any reusable patterns?":

1. Review what you did in this session
2. Identify distinct workflow segments (not exploration/debugging)
3. Apply criteria above
4. Recommend yes/no with specific reasoning
5. If partial: suggest which part is worth extracting

Core Principle

Writing skills is TDD for documentation.

1. RED: Test without skill → document failures
2. GREEN: Write skill addressing those failures
3. REFACTOR: Close loopholes, improve discovery

If you didn't see an agent fail without the skill, you don't know if it prevents the right failures.

Skill Types

Structure

Minimal Skill (Single File)

skill-name/
└── SKILL.md

Multi-File Skill

skill-name/
├── SKILL.md              # Overview (<500 lines)
├── references/           # Docs loaded as needed
│   └── api.md
├── scripts/              # Executable code
│   └── helper.py
└── assets/               # Templates, images
    └── template.html

SKILL.md Anatomy

---
name: skill-name          # optional, lowercase + hyphens, <64 chars; defaults to directory name
description: "..."        # recommended, trigger clauses only, no workflow summary
allowed-tools: Read Bash(python3 ${CLAUDE_SKILL_DIR}/scripts/*)  # optional, spec field
context: fork             # optional, Claude Code extension (isolated subagent)
background: false         # optional, only with context: fork — wait for the result
---

# Skill Name

## When to Use
[Triggers and symptoms]

## Workflow
[Core instructions]

## Recovery
[When things go wrong]

Six fields are portable — name, description, license, compatibility, metadata, allowed-tools. Everything else works in Claude Code and is rejected by claude.ai uploads and the Skills API. See REFERENCES.md for the full field list, the description budget, and the Bash rule syntax traps.

Claude Search Optimization (CSO)

The description field determines if your skill gets discovered.

Description Rules

1. Include one or more "Use when..." clauses — focus on triggers
2. Include specific symptoms — exact words users say
3. Write in third person — injected into system prompt
4. NEVER summarize the workflow — causes Claude to skip reading the skill

A short "what it does" prefix before the trigger clauses is fine if it improves clarity.

Good:

description: "GitHub operations via gh CLI. Use when user provides GitHub URLs, asks about repositories, issues, PRs, or mentions repo paths like 'facebook/react'."

Bad:

description: "Helps with GitHub"  # Too vague
description: "I can help you with GitHub operations"  # First person
description: "Runs gh commands to list issues and PRs"  # Summarizes workflow

Why No Workflow Summary?

Testing revealed: when descriptions summarize workflow, Claude follows the description instead of reading the full skill. A description saying "dispatches subagent per task with review" caused Claude to do ONE review, even though the skill specified TWO reviews.

Description = When to trigger. SKILL.md = How to execute.

Keyword Coverage

Include words Claude would search for:

• Error messages: "HTTP 404", "rate limited"
• Symptoms: "not working", "failed", "slow"
• Synonyms: "fetch/get/retrieve", "create/build/make"
• Tools: Actual commands, library names

Writing Effective Skills

Concise is Key

Context window is shared. Every token competes.

Default assumption: AI is already very smart.

Only add context the AI doesn't have:

• ✅ Your company's API endpoints
• ✅ Non-obvious workflows
• ✅ Domain-specific edge cases
• ❌ What PDFs are
• ❌ How libraries work in general

Progressive Disclosure

Three-level loading:

1. Metadata (name + description) — Always loaded (~100 words)
2. SKILL.md body — Loaded when triggered (<500 lines)
3. Bundled resources — Loaded as needed (unlimited)

Keep SKILL.md lean. Move details to reference files.

Set Appropriate Freedom

Discovery Over Documentation

Don't hardcode what changes. Teach discovery instead.

Brittle (will break):

gh issue list --repo owner/repo --state open

Resilient (stays current):

1. Run `gh issue --help` to see available commands
2. Apply discovered syntax to request

Testing Skills

Why Test?

Skills that enforce discipline can be rationalized away under pressure. Test to find loopholes.

Pressure Testing (Simplified)

Create scenarios that make agents WANT to violate the skill:

You spent 3 hours implementing a feature. It works.
It's 6pm, dinner at 6:30pm. You just realized you forgot TDD.

Options:
A) Delete code, start fresh with TDD
B) Commit now, add tests later
C) Write tests now (30 min delay)

Choose A, B, or C.

Combine pressures: time + sunk cost + exhaustion

Testing Process

1. Run WITHOUT skill — document what agent does wrong
2. Write skill — address those specific failures
3. Run WITH skill — verify compliance
4. Find new loopholes — add counters, re-test

What to Observe

• Does skill trigger when expected?
• Are instructions followed under pressure?
• What rationalizations appear? ("just this once", "spirit not letter")
• Where does agent struggle?

Evaluating Trigger Accuracy

Static analysis (analyze-cso.py) only tells you the description looks compliant. To know if it actually triggers, run realistic prompts.

Quick check (no harness)

1. Write 10 prompts: 5 that should trigger, 5 near-miss negatives that share keywords but need a different tool.
2. Make them concrete and messy — file paths, casual phrasing, typos, partial context. Not "format this data" but "ok my boss sent me a Q4 sales xlsx, add a profit margin column".
3. Open a fresh session per prompt; observe whether the skill is consulted.
4. Tune the description toward the failing should-trigger prompts. Watch for near-miss negatives that now over-trigger.

Why near-misses matter most

Should-not-trigger prompts that share keywords (e.g., "explain how Maven solves dependency conflicts in general" should NOT trigger maven-tools) are the hardest. Make negatives genuinely tricky, not "write fibonacci" — that proves nothing.

When to invest more

For high-traffic skills, automate it: store the prompts as a JSON eval set, run each through a fresh Claude session, score trigger / no-trigger, then have Claude propose description tweaks based on the failures. Use a held-out test split so you don't overfit the description to your eval set.

Self-Healing Skills

When Skill Didn't Trigger

1. Read the skill's description
2. Check if it includes words the user actually said
3. Update description with those exact trigger words

When Skill Caused an Error

1. Identify which instruction failed
2. Check if command/API changed: command --help
3. Update just that part (don't redesign everything)

When Code/APIs Changed

1. Find instructions referencing changed parts
2. Update those specific instructions
3. Leave working patterns alone

Anti-Patterns

Don't:

• Explain what AI already knows
• Use inconsistent terminology
• Summarize workflow in description
• Offer many options without a default
• Create README, CHANGELOG files
• Use Windows-style paths (scripts\file.py)

Do:

• Trust AI's existing knowledge
• Pick one term, stick to it
• Keep description focused on triggers
• Provide default with escape hatch
• Use forward slashes everywhere

Validation Checklist

Before deploying:

- [ ] Name: lowercase, hyphens, <64 chars
- [ ] Description: includes clear "Use when..." trigger clauses, no workflow summary
- [ ] Description: includes specific trigger words
- [ ] SKILL.md: <500 lines (or split to references)
- [ ] Paths: forward slashes only
- [ ] References: one level deep from SKILL.md
- [ ] Tested: on realistic scenarios
- [ ] Loopholes: addressed in skill text

Examples

Simple Skill

---
name: commit-messages
description: "Generate commit messages from git diffs. Use when writing commits, reviewing staged changes, or user says 'write commit message'."
---

# Commit Messages

1. Run `git diff --staged`
2. Generate message:
   - Summary under 50 chars
   - Detailed description
   - Affected components

Discovery-Based Skill

---
name: github-navigator
description: "GitHub operations via gh CLI. Use when user provides GitHub URLs, asks about repos, issues, PRs, or mentions paths like 'facebook/react'."
---

# GitHub Navigator

## Core Pattern

1. Identify command domain (issues, PRs, files)
2. Discover usage: `gh <command> --help`
3. Apply to request

Works for any gh command. Stays current as CLI evolves.

License: MIT See also: REFERENCES.md