AI-assisted codebase analysis methodology. Turns a complex codebase into a structured, browsable documentation set with clickable source links.
A reusable skill and toolkit for using AI agents (Claude Code, etc.) to systematically analyze large codebases. Born from the experience of analyzing vLLM — a ~250K line Python/C++/CUDA/Rust project — and producing the vllm-analysis documentation set.
- AI explores, human verifies — AI agents do the heavy research (finding files, reading code, extracting line numbers), but every claim is fact-checked before publishing
- Source links that never break — git submodule pins the analyzed version, all line references are permanent
- Local paths for editing, GitHub URLs for browsing — CI
sedstep converts links automatically - From general to specific — documents ordered so each builds on the previous
| File | Purpose |
|---|---|
| SKILL.md | The full methodology — what AI agents should read |
| templates/ | Templates for new analysis projects (README, config, CI) |
| scripts/ | Helper scripts (setup new repo, update submodule) |
- Read SKILL.md
- Copy
templates/to your new repo - Add the target project as a git submodule
- Start with the architecture overview, then work through each doc
- Fact-check before publishing
Analysis projects built with this methodology (alphabetical):
- aeron-analysis — Aeron: a high-performance, low-latency messaging transport for streaming data over UDP and IPC.
- ninja-analysis — Ninja: a small build system built with a singular focus on speed.
- upb-analysis — upb: a small, fast C implementation of Protocol Buffers.
- vllm-analysis — vLLM: a high-throughput, memory-efficient inference and serving engine for LLMs.
MIT