Feedback
I'm new to open source and wanted a first project I could actually finish. I started on a security testing tool, realized it was way too big, and pivoted to something every dev procrastinates on: writing docs.
docwrite reads a codebase and generates a README for it, using whatever Ollama model you already have installed. No cloud, no API keys, and it works offline. How it works: python/main.py auto-detects your installed Ollama model, reads the target project, sends it to the model and writes a README.md. Nothing is hardcoded to one model. It's plain Python and JS with no heavy frameworks, and MIT licensed. Testing so far: A small calculator project: accurate, usable README A multi-file Python + JS TODO app: coherent output with install, usage, features, contributing and license sections Both ran on 1.5B models (qwen2.5 and qwen2.5-coder) The limitation: small models sometimes invent plausible-but-wrong details. I've seen fake clone URLs and requirements.txt or npm scripts that don't exist. I'm documenting this in the project's README rather than hiding it. Where I'd love your input: Has anyone found a good way to stop small models from hallucinating setup steps? (Constraining output to only files that exist? Post-validation?) What would you want a README generator to get right first? Any advice for a first-time maintainer? Repo: https://github.com/sengoraku/docwrite Next I'm planning a supervised agent loop for local models with hard-coded safety rules, but I wanted to get docwrite right first.
- Reactions
- 0
- Comments
- 0
- Shares
- 0