One-shot a CLI coding agent on top of a Python LLM library
Simon Willison had the Fable model more or less one-shot a CLI coding agent built on his LLM Python library.
Done with Claude Code
What they did
Willison started a repo from his Python library template and gave Claude Code for web two prompts. The first wrote a spec.md for a Claude Code-style agent on the LLM alpha. The second told it to build the agent with red/green TDD in sensible commits, with docs, and to test manually using the OpenAI key in its environment.
How it went
It shipped as an alpha to PyPI and he called it pretty good for a first attempt. It added file and command tools and an unrequested Python API. A demo with GPT-5.5 built an ASCII-art clock app.
Worth knowing
The demo used the --yolo flag, which skips approvals. The README also shows --allow patterns, such as pytest*, as a narrower option.
Try it yourself with Claude Code
Using [Python LLM library, e.g. Simon Willison's llm] as a base, build a CLI coding agent that can read files, edit files and run shell commands in the current directory. Keep it to a single script with a clear README and a confirmation prompt before any command that writes or deletes. Finish by running it on a small demo task in [scratch folder] and showing me the transcript.
Source
Read the original