Choosing the Right Model
Choosing the right model and iteration count is the single biggest factor affecting output quality. This guide provides battle-tested recommendations based on real-world usage.
Decision Matrix
| Project Type | Example | Recommended Model | Iterations (-i) |
Est. Cost | Est. Time |
|---|---|---|---|---|---|
| Small library (<20 files) | pallets/click, lodash |
openai/gpt-4o |
1 |
~$0.05 | <30s |
| Medium application (20-100 files) | expressjs/express, fastify |
openai/gpt-4o |
5 |
~$0.20 | 1-2 min |
| Large application (100-500 files) | Single-stack web apps | openai/gpt-4o |
10-15 |
~$0.50 | 3-5 min |
| Monorepo (multi-stack) | React + Python + APIs | openai/gpt-5.2 |
25 |
~$1.50 | 5-10 min |
| Enterprise codebase (500+ files) | Microservices, large orgs | anthropic/claude-sonnet-4.6 |
35 |
~$3.00 | 10-15 min |
Why These Recommendations?
Small Libraries (-i 1)
For projects with fewer than 20 files, the entire source tree fits comfortably in a single LLM context window. One iteration is enough because the AI can see everything at once.
npx @thelogicatelier/sylva --github-repository https://github.com/pallets/click -m openai/gpt-4o -i 1
Medium Applications (-i 5)
At 20-100 files, the RLM agent needs a few passes to traverse nested directories and cross-reference imports with dependency files. Five iterations lets it scan the top-level structure, then drill into 3-4 key subdirectories.
npx @thelogicatelier/sylva --local-repository ./my-express-app -m openai/gpt-4o -i 5
Large Monorepos (-i 25)
This is where model choice matters most. Multi-stack repos (e.g., React frontend + Python FastAPI backend + Wix API integration) require:
- Many iterations to traverse both the frontend and backend subtrees
- A strong reasoning model to avoid hallucinating frameworks
Real-world example: myshabeauty — a React/Tailwind frontend with a Python FastAPI backend, Wix API integration, and Fly.io deployment:
npx @thelogicatelier/sylva --local-repository . -m openai/gpt-5.2 -i 25
This produced a detailed AGENTS.md correctly identifying:
- React + Tailwind CSS + CRACO (frontend)
- Python + FastAPI (backend)
- Wix APIs + Instagram Graph API (integrations)
- Docker + Fly.io (deployment)
- Pytest (testing)
- Context API for state management
Enterprise Codebases (-i 35)
The default iteration count of 35 is designed for the largest repositories. Use Anthropic's Claude Sonnet for its massive context window and strong instruction-following.
npx @thelogicatelier/sylva --local-repository ./enterprise-monorepo -m anthropic/claude-sonnet-4.6 -i 35
Model Comparison: When to Use What
OpenAI gpt-4o — The Workhorse
- Best for: Most projects, predictable costs
- Strengths: Fast, reliable, good balance of accuracy and cost
- Weaknesses: May need more iterations for very large repos
OpenAI gpt-5.2 — The Powerhouse
- Best for: Complex monorepos, multi-stack architectures
- Strengths: Superior reasoning, handles cross-stack detection well
- Weaknesses: Higher cost per token
Anthropic claude-sonnet-4.6 — The Deep Thinker
- Best for: Enterprise codebases, maximum accuracy requirements
- Strengths: Massive context window, excellent instruction-following
- Weaknesses: Slower than GPT models
Google gemini-3.1-pro — The Default
- Best for: Quick analysis, cost-sensitive scenarios
- Strengths: Generous free tier, good for experimentation
- Weaknesses: May be less precise on complex architectural patterns
Cost Optimization Tips
- Start small: Run with
-i 1first to check basic output, then increase - Use the right model tier: Don't use
gpt-5.2on a 5-file project - Iterate on prompts, not iterations: If output is wrong, the issue may be in the prompts rather than needing more iterations
- Cache results: Generated
AGENTS.mdfiles don't expire — only regenerate when the codebase significantly changes