Skill Development

Skill Development Guidelines

Before You Start

Define 2-3 concrete use cases before writing:

1. What does a user want to accomplish?
2. What multi-step workflows does this require?
3. What domain knowledge should be embedded?

Scale awareness: This repo has 30+ active skills. Every new skill adds ~100 tokens of description scanning to every prompt. Before creating a new skill, check if an existing skill can be extended instead.

Skill Categories

See references/workflow-patterns.md for 5 common skill patterns.

Skill Durability

Skills also differ in how long they remain useful:

Most skills in this repo are Encoded Preference (global conventions). See Skill Lifecycle for when to deprecate Capability Uplift skills. See references/workflow-patterns.md for the Encoded Preference vs Capability Uplift distinction in detail.

Folder Structure

skill-name/           # kebab-case required
├── SKILL.md          # Required (uppercase, exact)
├── references/       # Detailed docs (on-demand loading)
├── scripts/          # Executable scripts
└── assets/           # Templates, fonts, icons

Critical Rules:

• SKILL.md must be exactly SKILL.md (case-sensitive)
• Folder name: kebab-case, no spaces, no capitals
• NO README.md inside skill folder (use references/)

YAML Frontmatter

Required Fields

---
name: skill-name      # kebab-case, matches folder
description: "..."    # Controls when skill triggers
---

Optional Fields

license: MIT
compatibility: Requires Python 3.10+
metadata:
  author: Your Name
  version: 1.0.0
  mcp-server: server-name

See references/yaml-reference.md for complete field documentation.

Security Restrictions

Forbidden in frontmatter:

• XML angle brackets (< >) - injection risk
• "claude" or "anthropic" in name (reserved)

Description Field (CRITICAL)

Structure: [What it does] + [When to use it] + [Key capabilities]

# BAD: Too vague
description: "Creates documents"

# BAD: Missing triggers
description: "Sophisticated multi-page documentation systems"

# GOOD: Complete formula
description: "Creates Word documents with tracked changes for legal review. Use when user asks for contracts, legal documents, or revision-tracked content."

# GOOD: With negative trigger
description: "Advanced data analysis for CSV files. Use for statistical modeling, regression. Do NOT use for simple data exploration."

Include:

• What the skill does
• Trigger phrases/scenarios (what users actually say)
• File types if relevant
• Negative triggers when needed (Do NOT use for...)

Description Analysis

With 30+ skills loaded, false positives (skill triggers when it shouldn't) waste context, and false negatives (skill doesn't trigger when needed) lose value. Analyze descriptions systematically:

False Positive Check: List 5 prompts from adjacent domains. Does your skill trigger? If yes, add negative triggers.

False Negative Check: List 5 paraphrased prompts for your skill. Would the description match? If not, add trigger phrases.

Overlap Check: Search existing skill descriptions for keyword collisions. If two skills trigger on the same prompt, either:

• Narrow one with negative triggers
• Merge if they cover the same domain

Content Guidelines

Litmus Test

"If Claude did NOT read this SKILL, would its output quality drop?"

• Yes --> keep the content
• No --> remove it (Claude already knows this)

DO Include

• Policies/directions: What to do, what approach to take
• Anti-patterns: What to avoid — but ONLY if they pass the Litmus Test above. If Claude would avoid the pattern without reading the skill, do not include it.
• Domain knowledge: Information requiring lookup
• Output formats: Expected structure when not obvious
• Review categories: Platform-agnostic area names that anchor coverage (see Review Categories Pattern below)

DO NOT Include

• Standard library commands (Claude knows these)
• Basic language syntax
• General programming concepts
• Things you could find in any tutorial
• Scope mismatch: Platform-specific content in a skill whose agent targets all platforms (e.g., React checklist in a general performance-reviewer). Platform-specific content is correct when the skill's scope matches its agent's scope
• Context-dependent severity thresholds: The same pattern (e.g., O(n^2)) can be CRITICAL or harmless depending on data size and context -- do not hardcode severity in cross-platform skills

Simplification Safety Checklist

When removing content from skills, verify:

• WRONG/CORRECT code contrasts: Keep for style-consistency skills (coding-style, etc.) — without concrete examples, judgment drifts across runs
• Cross-skill references: "See makefile-first" or "See /tdd" sections connect skill dependencies — removing them breaks agent routing
• Reference file targets: If text says "See references/X.md", verify X.md actually contains the promised content
• Agent dependencies: Check skills: frontmatter in all agents that load this skill — don't remove content they depend on
• Coverage targets: Language-specific metrics encode genuinely different policies — do not consolidate even if they look similar

Calibration Examples

Review Categories Pattern

Review-oriented skills and agents need anchors to ensure reproducible coverage across runs. Without any structure, Claude's non-deterministic output means reviews can vary -- one run checks caching, the next skips it entirely.

Solution: Use categories, not checklist items.

Categories go in the agent (not a skill) when they define how the agent structures its review workflow. Categories go in a skill when they define shared principles multiple agents reference.

Progressive Disclosure

Skills load in three levels to save context:

Front-load important rules - Claude pays more attention to content at the top.

Size Guidelines

Testing Your Skill

See references/troubleshooting.md for common issues (trigger failures, instructions not followed).

Iteration Signals

See references/troubleshooting.md for common issues.

Structured Evals

For systematic skill quality measurement, see references/eval-patterns.md.

Skill Lifecycle

When to Deprecate

Review skills when:

• A new model release may have learned the skill's techniques natively
• The skill's pass rate in evals drops below the no-skill baseline
• Team conventions change, making the skill's policies outdated

Deprecation Process

1. Run the skill's eval prompts WITHOUT the skill loaded
2. Compare output quality against WITH-skill results
3. If no-skill output is equivalent or better: delete the skill folder
4. Update any agents that referenced it in skills: frontmatter

Checklist

Before Creating

• Identified 2-3 concrete use cases
• Is this knowledge Claude doesn't have?
• Would I explain this repeatedly without a skill?

During Development

• Folder name in kebab-case
• SKILL.md exists (exact spelling)
• Frontmatter has --- delimiters
• Name field: kebab-case, no spaces
• Description includes WHAT and WHEN
• No XML tags anywhere
• Under 500 lines
• Anti-patterns included only if they pass the Litmus Test (Claude wouldn't know without the skill)
• Library-specific: verify with context7 if APIs may have changed. Differences from Claude's knowledge → write in SKILL.md

Before Release

• Tested triggering on obvious tasks
• Tested triggering on paraphrased requests
• Verified doesn't trigger on unrelated topics
• Functional tests pass

Ongoing Maintenance

• Classified as Encoded Preference or Capability Uplift
• Description analyzed for false positives/negatives against adjacent skills
• Capability Uplift skills: reassessed after model updates