Writing Documentation

Writing Documentation Skill

Variables

Commands

• CLAUDE_PRINT_COMMAND: claude --print --model haiku

Agents

• RESEARCHER_AGENT: researcher
• DOCS_REVIEWER_AGENT: docs-reviewer

Documentation Types (Diátaxis Framework)

• Tutorial - Learning-oriented, step-by-step
• How-to - Task-oriented, specific problems
• Reference - Technical specifications
• Explanation - Clarifies concepts

Details in references/best-practices.md.

Writing for Claude Code

CRITICAL: When writing documentation that Claude reads (SKILL.md, CLAUDE.md, commands, agents):

1. Test Claude's Base Knowledge First

Verify what Claude already knows:

# Use haiku for cost-effective testing (gives same quality answers as sonnet)
$CLAUDE_PRINT_COMMAND "Do NOT use any skills. How would you [perform task]?"
$CLAUDE_PRINT_COMMAND "Do NOT use any skills. When should you [make decision]?"

2. Document ONLY Unique Patterns

Include only what Claude wouldn't naturally do:

• ✓ Opinionated architectural choices
• ✓ Counter-intuitive decisions
• ✓ Project-specific conventions
• ✓ Non-default patterns

Remove redundant content:

• ✗ Standard library usage
• ✗ Common best practices
• ✗ Well-known patterns
• ✗ Basic language features

3. Example: React Skill Reduction

Testing revealed Claude knows TanStack Query/Zustand/RTL patterns but doesn't default to:

• "Test stores, not components" (counter-cultural)
• "NO useState for complex logic" (prescriptive)
• "Inline actions unless repeated 2+" (specific rule)

Result: 328→125 lines (-62%) by documenting only unique opinions.

Verifying Technical Accuracy

API Verification Workflow

When documenting unfamiliar APIs or libraries:

1. Launch $RESEARCHER_AGENT:

Use Task tool to launch `$RESEARCHER_AGENT` agent to verify [API/library] documentation

$RESEARCHER_AGENT uses Context7 MCP to fetch official API docs and verify method signatures.

2. Read the codebase:

For internal/project APIs:

Read relevant source files to verify method signatures exist

3. State version requirements:

• Specify versions when certain: # Using pandas 2.0+ DataFrame.merge()
• Add verification note when uncertain: # Verify this API exists in your version

4. Direct to official docs: Add link to authoritative source.

Security Verification

Required checks before documenting code:

1. SQL: Parameterized queries, never string concatenation
2. YAML: yaml.safe_load(), never yaml.load()
3. Credentials: Environment variables, never hard-coded
4. Input: Always validate before processing
5. Errors: Handle network/file operations

Use $RESEARCHER_AGENT agent if uncertain about security best practices.

Code Example Requirements

Every Example Must Include

1. All imports and dependencies
2. Complete, copy-paste ready code (no ellipsis or pseudo-code)
3. Expected output when relevant
4. Error handling for production use
5. Context explaining "why"

Example:

# Process in batches of 1000 to avoid memory exhaustion.
# Testing: smaller (100) = 3x overhead, larger (10000) = OOM on 8GB systems.
BATCH_SIZE = 1000

for batch in chunks(items, BATCH_SIZE):
    process_batch(batch)

Production-Ready Requirements

Include when relevant:

• Authentication/authorization
• Logging for debugging
• Rate limiting and retries
• Timeout handling
• Resource cleanup

See references/best-practices.md for complete production-ready examples.

Using $DOCS_REVIEWER_AGENT

After writing documentation:

Use `$DOCS_REVIEWER_AGENT` agent to ruthlessly simplify

The agent challenges every element's necessity, asking "Would the documentation still be clear without this?"

Critical for:

• CLAUDE.md, SKILL.md files
• Slash commands and agents
• Any Claude Code documentation

Eliminates:

• Motivational language
• Redundant examples
• Unnecessary context that doesn't change behavior

LLM Self-Checklist

Before publishing:

Verification:

• Verified APIs exist ($RESEARCHER_AGENT, codebase read, or official docs)
• Code is complete and secure (all imports, parameterized queries, error handling)
• Examples are production-ready (auth, logging, retries, timeouts)

Content Quality:

• Context and "why" explained, not just "what"
• Specific details, not generic (40ms not "significant", name technologies not "various aspects")
• Consistent terminology throughout
• Appropriate hedging (direct for facts, hedge only when uncertain)

Claude Code Docs:

• Tested base knowledge with $CLAUDE_PRINT_COMMAND
• Documented only unique patterns
• Applied $DOCS_REVIEWER_AGENT agent for ruthless simplification

References

Research foundation in references/:

• exemplary-projects.md - Analysis of well-documented projects
• best-practices.md - Authoritative style guide synthesis
• llm-pitfalls.md - LLM-specific quality issues