Claude Code builds a Python coding-agent package
Simon Willison used Claude Code for web to specify and implement llm-coding-agent, published an alpha package and tried its file and command tools on a small Swift task.
simonwillison.net
Original prompt
Shared by the source author. See source notes for context.
Write a spec.md for this project
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Simon Willison
- Original account
- Read the source ↗
- Published
- 2 Jul 2026
- Source link checked
- 15 Oct 2026
- Reported result
- Willison published llm-coding-agent 0.1a0 with file and shell tools, linked its development artifacts and reported a working initial trial.
- Human involvement
- Willison supplied two high-level requests, inspected the output, published the alpha and manually tried the new package.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Start a Python package repository and request an implementation specification.
- Ask Claude Code to implement tools in commits with tests and documentation.
- Publish the alpha package and inspect its tool inventory and session artifacts.
Simon Willison wanted to see what a coding agent built on his Python LLM library would look like. He started a repository from his package template and asked Claude Code for web to write a specification for a command-line assistant with file editing and shell tools.
A second instruction told the agent to commit the specification, implement it in staged commits, keep tests and documentation current and occasionally try the implementation against an available OpenAI API key. Willison links the session transcript, specification, README and commit history. He then published the resulting alpha on PyPI.
The package exposed reading, writing, searching, targeted edits and command execution. It also included a Python interface Willison had not requested. In his trial, the new agent handled a request for an ASCII clock application and recognized that SwiftUI was unsuitable for a true terminal program.
The evidenced deployment is Claude Code building a usable experimental package. The later clock trial used the new package with GPT-5.5 and should not be attributed to Claude. Willison calls the release an early alpha; the account does not establish production reliability or an independent security review.
Prompt
Original prompt
Source: Blog · Simon Willison
Frequently asked questions
Which agent built the package?
The author links a Claude Code for web transcript for specification and implementation.
Did Claude also run the ASCII clock trial?
The source identifies GPT-5.5 reasoning inside the newly built package for that later trial.
Was the result production-ready?
Willison calls it a slop-alpha and reports a first attempt, without a production reliability assessment.
More details
About this account
Claude Code writes the specification, implements file/shell tools in staged commits and updates tests and documentation.
Firsthand primary account opened and read; original publication date checked in primary metadata or visible timestamp. Outcomes are author-reported, not independently reproduced. The downstream package trial used GPT-5.5; alpha release and checks are author-reported. Original prompt excerpt reproduced exactly from the source; excerpt only, not the complete prompt.
Credit: Simon Willison
Source published 2 Jul 2026
Source checked 15 Oct 2026
Tools and topics
Claude Code for web · Python · LLM library · Git · PyPI
Software · Reported deployment
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.