Building a 350K-Line Codebase Solo in 52 Days with AI Agents
The Codebase Is the Context
A 350,000-line codebase in 52 days sounds impossible for a solo developer. But the math changes when your AI agents have the right context. The codebase itself becomes the specification - the agents read existing patterns, follow established conventions, and extend what already exists.
The key insight: most of your time goes into context, not code.
CLAUDE.md Files Are Your Multiplier
Every project directory gets a CLAUDE.md file that tells the agent how to work in that part of the codebase. What conventions to follow. What patterns to use. What mistakes to avoid.
This is not documentation for humans. It is executable context for agents. When an agent opens a directory and reads the CLAUDE.md, it immediately knows:
- The architecture pattern for this module
- Which libraries and APIs to use
- Error handling conventions
- Testing expectations
- Common pitfalls specific to this area
Without this, agents produce code that works but does not fit. With it, they produce code that looks like you wrote it.
The Daily Workflow
The actual workflow is surprisingly repetitive:
- Write the spec. Describe what you need in precise terms. This takes most of your time.
- Set agents loose. Multiple agents work in parallel on different parts of the codebase.
- Review and integrate. Check the output, fix integration issues, commit what works.
- Update context files. When you find patterns the agents get wrong, update the CLAUDE.md files so they do not repeat the mistake.
Step 4 is the compound interest. Every context update makes all future agent work better. By week 3, the agents rarely produce code that needs significant revision.
What This Means for Solo Developers
You are not writing 350K lines. You are writing specs and context files - maybe 10K lines of actual human-written content. The agents produce the rest. But you need deep understanding of the entire codebase to write good specs and catch bad output.
Solo AI-augmented development is not about working less. It is about working differently.
Fazm is an open source macOS AI agent. Open source on GitHub.