Claude Code

Claude Code is Anthropic's official CLI tool for AI-assisted development. In this documentation system, it serves as the primary orchestration layer.

Overview

Model Hierarchy

Model	Role	Use Case
Claude Opus	Orchestrator	Complex decisions, context merging, final quality control
Claude Sonnet	Worker Agent	Parallel source reading, chapter writing, code generation

Installation

# Install Claude Code CLI
npm install -g @anthropic-ai/claude-code

# Authenticate
claude auth login

# Verify installation
claude --version

Project Configuration

Create .claude/CLAUDE.md in your project root:

# Claude Code Project Configuration

## Project Overview
This is a documentation generation project using LaTeX.

## Key Files
- `latex/main.tex` - Main document
- `latex/bibliography.bib` - References
- `latex/chapters/` - Chapter files
- `agents/AGENT_GUIDELINES.md` - Writing guidelines

## Agent Usage

### For Chapter Writing
Use **Opus** model for complex chapters:
- Max pages: X
- Sources: [list]
- Style: flowing prose, not bullets

### For Quick Tasks
Use **Sonnet** for:
- Bibliography fixes
- Simple corrections
- Code examples

Orchestration Pattern

Spawning Parallel Agents

The orchestrator (Opus) can spawn multiple Sonnet agents to work simultaneously:

User: "Analyze all source documents"

Opus (thinking):
- I have 3 PDFs in sources/
- I'll spawn 3 agents to read them in parallel
- Each agent will return a summary
- I'll merge the summaries

Opus → Task: "Read sources/thesis1.pdf and summarize" (Sonnet)
     → Task: "Read sources/thesis2.pdf and summarize" (Sonnet)
     → Task: "Read sources/report.pdf and summarize" (Sonnet)

[Agents work in parallel]

Opus receives: Summary1, Summary2, Summary3
Opus merges: Combined context for next phase

Agent Task Invocation

When the orchestrator delegates a task:

Task agent invocation:
├── Agent type: Sonnet (efficient for well-defined tasks)
├── Context: Relevant source excerpts only
├── Instructions: Specific chapter requirements
├── Constraints: Page limit, style guidelines
└── Output format: Complete LaTeX chapter

Best Practices

1. Context Management

Keep Context Focused

Don't pass entire documents to worker agents. Extract only relevant sections to keep context windows efficient.

Good:
- "Read pages 10-25 of thesis.pdf about security threats"

Bad:
- "Read the entire thesis.pdf"

2. Task Granularity

Break large tasks into parallelizable units:

Instead of:
- "Write the entire document"

Use:
- "Write chapter 1: Introduction (4 pages)"
- "Write chapter 2: Background (8 pages)"
- "Write chapter 3: Methodology (10 pages)"
[Parallel execution]

3. Quality Checkpoints

Build in review stages:

Configuration Files

.claude/settings.json

{
  "model": "claude-opus-4-20250514",
  "agentModel": "claude-sonnet-4-20250514",
  "maxTokens": 8192,
  "temperature": 0.7
}

Project Context File

The .claude/CLAUDE.md file provides project context to Claude Code:

# Project: AI Security Documentation

## Goal
Generate 150-page research document on AI security.

## Source Files
- sources/thesis1.pdf - ML attack research
- sources/thesis2.pdf - Defense mechanisms
- sources/report.pdf - Industry analysis

## Output
- latex/chapters/*.tex - Chapter files
- latex/main.tex - Main document

## Style Guidelines
- Academic prose, not bullet lists
- Proper language with correct accents
- Max 2-3 code blocks per chapter
- Citations using \cite{}

Workflow Commands

Starting a Session

# Start Claude Code in project directory
cd your-project/
claude

# Claude will read CLAUDE.md and understand project context

Common Operations

# Ask Claude to analyze sources
> Analyze all documents in sources/ and create a synthesis

# Generate a chapter
> Write chapter 05 following the chapter-writer prompt

# Review a chapter
> Review latex/chapters/05-security.tex for quality

Integration with External Tools

Claude Code can generate prompts for external AI services:

The orchestrator generates research prompts, you run them on external platforms, then paste results back for integration.

Error Handling

Common Issues

Error	Solution
Context too long	Split into smaller tasks
Rate limit exceeded	Wait or use lower tier
Agent timeout	Reduce task complexity

Recovery Pattern

If agent fails:
1. Log the error
2. Reduce scope
3. Retry with simpler task
4. Escalate to orchestrator if repeated failures