How do I write a spec a Claude Code crew can build small features from in a Python backend?

Tares writes the spec with you. Run tares factory spec with a one-line idea in your Python repo, and /tares:spec interviews you one question at a time and has a second model attack the plan. It then writes a Tares project with the spec, milestones with acceptance checks, tickets with a working doc each, and an AGENTS.md. You answer questions instead of drafting a document.

How it works

/tares:spec runs inside Claude Code, in the repo the project builds on or in a folder that holds its repos. Here is what happens, in order:

• The interview. It asks one question at a time. Your answers are the raw material for the plan, so be concrete about the feature you want built.
• The attack. A second model attacks the plan before anything is written, so weak points surface while they are still cheap to change.
• The write-up. The plan becomes a Tares project containing the spec, milestones with acceptance checks, tickets with a working doc each, and an AGENTS.md.
• The hand-over. It ends by asking whether to give the project to the crew. If you say yes, the orchestrator picks it up.

The crew then works from that record. The orchestrator hands out work. Builders, one per milestone and repo, each take a ticket, work on their own branch and open a pull request. The reviewer holds every pull request to a fixed bar, and a builder merges only after the reviewer passes it and CI is green. Tares keeps the plan, tickets, reviews, the assumptions the crew made and your answers, so a crew that restarts carries on where it stopped.

For a Python backend, tares factory doctor reads each handed-over repo’s languages from manifests such as pyproject.toml, setup.py, setup.cfg, Pipfile or requirements.txt. It tells you whether Claude Code has a language server for Python and prints what to run when one is missing. Builders and the reviewer search less when a server can find definitions, references and callers. tares factory crew up offers the same install and runs it only when you say yes.

Set it up

1. Check that Claude Code is signed in, gh is signed in with write access to the repos the crew works in (gh auth login), and git, jq and curl are installed.
2. Install the CLI and start Tares locally. On Tares Cloud, run tares login instead and pick your workspace in the browser.
3. In a second terminal on a local Tares, start the MCP endpoint every crew session shares. The crew reaches Tares only through it.
4. Give Tares access to GitHub: Settings > GitHub on a local Tares, or Connect GitHub on Tares Cloud.
5. Bring up the crew. The first run asks how far the crew may go, for example whether it merges on its own after review and green CI or waits for your yes.
6. Plan the feature from your Python repo, with the idea in a sentence or two.

uv tool install tares     # or: brew install glassflow/tap/tares
tares up
tares mcp --transport streamable-http
tares factory crew up
tares factory spec "what you want to build"

7. Answer each interview question in the Claude Code session that opens. Give specifics: the behavior you want, how you will know it works, and anything the plan must not touch. Those answers shape the acceptance checks the crew builds against.
8. When the plan is written, say yes to the hand-over. Questions from the crew then reach you in the orchestrator’s session, one at a time, each with a recommendation.

tares factory crew attach orchestrator

If your tickets should live in GitHub Issues, say so during /tares:spec and name the repo. Claude creates the issues with gh and links them. See Get started for the full setup, including Linear.

What you get

• A Tares project with a spec, milestones with acceptance checks, and tickets that each carry a working doc.
• An AGENTS.md written as part of the plan, and your shared AGENTS.md applies to sessions that work on the project.
• Builders that each take one ticket onto its own branch and open a pull request, with the reviewer posting its verdict on every one.
• A record on Tares of the assumptions the crew made, the questions it asked and your answers in your own words. You can check the crew’s own choices at the end of each milestone instead of during every step.
• Progress on the Crew page of the Tares console, with tares factory crew status and tares factory projects on the command line.

When something else fits

• If the work is small enough for one session, run tares factory session with your idea. It asks a few questions, writes a short spec and a ticket list on Tares, and builds it right there. If the plan turns out big, it offers the full spec and the hand-over to the crew.
• If a station misbehaves or the setup looks wrong, run tares factory doctor. It prints each problem with the command that fixes it.

Related