Orchestrate AI agents to transform sources into professional LaTeX documents
The documentation generation system uses a hierarchical multi-agent architecture where Claude Opus serves as the orchestrator, delegating tasks to specialized Sonnet agents that work in parallel.
| Tool | Role | Model |
|---|---|---|
| Claude Code | Primary orchestration & writing | Opus (orchestrator), Sonnet (agents) |
| OpenCode | Alternative multi-provider interface | Configurable |
| Service | Purpose | Output |
|---|---|---|
| NotebookLM | Source document synthesis | Structured summaries, cross-references |
| Perplexity | Current statistics & trends | Market data, recent developments |
| Scopus AI | Academic paper search | BibTeX citations, peer-reviewed sources |
| Component | Technology |
|---|---|
| Document Format | LaTeX with BibLaTeX |
| Build System | pdflatex + biber |
| Version Control | Git |
| Documentation | VitePress |
Model: Claude Opus
Role: Controls workflow, merges context, delegates tasks, makes decisions
Tasks:
Model: Claude Sonnet
Role: Execute parallel tasks efficiently
Tasks:
Model: Claude Sonnet/Opus
Role: Quality assurance & feedback
Tasks:
# Clone template
git clone https://github.com/your-repo/ai-doc-template.git
cd ai-doc-template
# Verify installation
./scripts/verify-setup.sh# Copy source documents
cp ~/research/*.pdf sources/
# Sources will be analyzed by parallel agents# Launch Claude Code with Opus
claude
# The orchestrator will guide the process# Compile LaTeX to PDF
./scripts/compile.sh
# Output: latex/document.pdfCollect all source documents into the sources/ directory. The system supports:
The orchestrator spawns multiple Sonnet agents to read sources simultaneously:
Opus → spawn Agent1 → read sources/thesis1.pdf
→ spawn Agent2 → read sources/thesis2.pdf
→ spawn Agent3 → read sources/report.docx
[All agents work in parallel]
Agent1 → summary → Opus
Agent2 → summary → Opus (Context Merge)
Agent3 → summary → OpusBased on identified gaps, the orchestrator generates prompts for external AI tools:
| Tool | Input | Output |
|---|---|---|
| NotebookLM | Uploaded source PDFs | Cross-referenced synthesis |
| Perplexity | Generated research prompt | Current statistics, trends |
| Scopus AI | Academic search queries | BibTeX citations |
With full context, Opus creates:
Writer agents receive specific assignments and work simultaneously:
Chapter assignments:
├── Agent1 → 01-introduction.tex (4 pages)
├── Agent2 → 02-background.tex (8 pages)
├── Agent3 → 03-methodology.tex (10 pages)
├── Agent4 → 04-implementation.tex (12 pages)
└── Agent5 → 05-results.tex (8 pages)
Each agent receives:
- Relevant source excerpts
- Research data
- Style guidelines
- Page budget
- Chapter templateThe review loop ensures quality: