Claude Agents Python — Testing skill for Claude Code

claude-agents-python

Seven role-scoped Claude Code subagents for solo Python development. Write permissions are deliberately split by role — the implementer can't touch tests, the test author can't touch source — so you stay the orchestrator.

[](./LICENSE)

What this is

A set of Markdown files you drop into `.claude/agents/`. Each defines a specialist with its own system prompt, tool allowlist, and model. Claude delegates to them; you decide what gets built and which findings matter.

Solo development has no second pair of eyes and no division of labor. A single agent doing everything will, sooner or later, weaken a test to make it pass or refactor something you never asked about. Splitting the work into roles with different permissions makes those failures structurally impossible rather than merely discouraged.

The design: asymmetric write permissions

This is the part that matters more than the role names.

Agent	Can write to	Model
librarian	nothing	sonnet
implementer	source only	sonnet
test-author	tests/ only	sonnet
test-runner	nothing	haiku
debugger	anywhere	opus
reviewer	nothing	opus
documenter	docs and docstrings	sonnet

`implementer` cannot edit `tests/`, so it can never make a failing test pass by rewriting the test. `test-author` cannot edit source, so it can never bend a test to match a buggy implementation. `reviewer` and `test-runner` carry `disallowedTools: Write, Edit`, so critique stays separate from correction.

When one agent holds every permission, these failures happen quietly. Separating them forces every real decision back to you.

Model assignment follows the same logic: opus where judgment drives the outcome, haiku where the job is summarizing output.

Quick start

git clone https://github.com//claude-agents-python.git
cd claude-agents-python

# Opti