// Before you install anything

Try provisioning it

Not a recording. You type the command, you answer the real coldstart interview — the same questions, defaults and choices the engine ships — and at the end the actual generator runs on your answers and hands you the two files it produces, to read here or download — including what it writes when you ask for an autonomy dial above its ceiling.

Three of the four things on this page are the engine’s, not ours: the dry-run output is captured from scripts/coldstart.sh --dry-run, the questions are read from its questions.json, and the files at the end come from generate.py run server-side on what you typed. The notes in the margin are ours, and say so.

Loading the walkthrough…

The same thing, in words

For anyone reading without JavaScript, or who would rather not type: coldstart.sh --dry-run prints an ordered plan and exits before touching anything. It halts at step 7 — HALT — and that halt is deliberate: provisioning wires everything mechanical and then stops, because the next decision is what the team should actually work on, and a script guessing at your backlog would be worse than stopping.

Past the halt comes the coldstart interview: 19 core questions about your project’s identity, stack, deploy target and autonomy dials. Your answers go to scripts/coldstart-interview/generate.py, which writes two files — config.json (the dial defaults and the active role roster) and CLAUDE.project.md (a project overlay, not the canonical CLAUDE.md).

The dials have ceilings. Ask for external.system at level 5 and the generated file says 2, because that is the ceiling and no answer raises it. There is no override flag; the number in the file is simply the smaller one. That is the mechanism the safety page describes.

Then what?

The plan above stops at step 7 on purpose. Provisioning wires everything mechanical and then halts, because the next decision — what should the team actually work on — is not one a script should guess at. You fill in an epic, re-run with --resume, and it seeds that backlog as GitHub Discussions.

From there every change follows the same path, and you can watch a real one go through it: one pull request, end to end — the objections, the re-reviews, the gates and the merge.