Sign in
Coming soon

The repo agent

Point it at a repository. It reads the code, the tests and the CI, then writes back a design and a Claude Code setup made for that code, with what it checked and what it didn't.

Start with the lessonsExplore the catalog

How it will work

Preview, coming soon: an illustration of the repo assistant, which is not available yet. A repository link is pasted, the assistant reads 214 files, ranks three skills, and adds the best match, test-driven-development, to the repository's .claude/skills folder.

  1. Reads the repository: the package, the tests and the CI
  2. Finds the commands that build, lint and test it
  3. Reads the agent rules already there (AGENTS.md, skills)
  4. Writes CLAUDE.md, two skills, a reviewer subagent, settings and a hook
  5. Lists what it checked, and what you should run yourself

A demo. The repo agent isn't live yet.

Example report: github/spec-kit

Written by hand from the repository's own files, to show what the repo agent will produce.

The repository

github/spec-kit at commit c00dc05: the Specify CLI (specify), a Python 3.11+ command-line tool built with Typer and packaged with Hatchling, plus the prompt templates, scripts, presets and extensions it installs into other projects.

What we found Where
Python package, about 240 files src/specify_cli/
Tests with pytest, about 320 files tests/ (specify_cli, integrations, extensions, workflows, contract, hooks)
Lint with Ruff 0.15.0, tests on Linux, Windows and macOS with Python 3.13 and 3.14 .github/workflows/test.yml
Shell scripts checked with ShellCheck .github/workflows/lint.yml
Agent instructions already present AGENTS.md (58 lines), .github/skills/code-review, .github/skills/add-community-extension
Design notes the agent must read before certain changes design/cli.md, design/integration.md, design/workflow-step.md

What a coding agent needs here

  1. The right test command. AGENTS.md warns that a bare uv run pytest can pick up an editable install from another worktree. The safe command is uv sync --extra test, then .venv/bin/python -m pytest.
  2. Read-before-change rules. CLI commands, agent integrations and workflow steps each have a design note to read first. An agent that skips them breaks the structure the maintainers rely on.
  3. Disclosure rules. CONTRIBUTING.md asks for disclosure on agent-authored commits, pull requests and review comments, each separately.
  4. Evidence. Behavior changes need positive and negative tests; bug fixes need before-and-after evidence (CONTRIBUTING.md, "Testing deterministic behavior").

The proposed Claude Code setup

File Kind What it does
CLAUDE.md Rules Imports AGENTS.md (@AGENTS.md) so both agents follow one source, and adds the test command and the three design notes as a short checklist
.claude/skills/run-tests/SKILL.md Skill Syncs the worktree's own environment and runs Ruff and pytest the way CI does, then reports failures with file and line
.claude/skills/design-check/SKILL.md Skill Given a change, names the design note it touches (cli, integration or workflow-step) and lists the rules from it that apply
.claude/agents/test-reviewer.md Subagent Reviews a diff for missing negative tests and missing before-and-after evidence, read-only
.claude/settings.json Settings Allows uv sync, uv run, .venv/bin/python -m pytest and uvx ruff; asks before git push
.claude/hooks/ruff-on-edit.sh Hook After a Python file is edited, runs uvx ruff@0.15.0 check on that file and shows the result

What was checked, and what wasn't

  • Checked by hand: every path and command above exists at commit c00dc05.
  • Not run: the hook and the skills were not executed. Read them before installing, and run the test skill once yourself.

This example was written by hand to show what the repo agent will produce. The repo agent itself is coming soon.