Set Up Brain

Build a small workspace around the user's actual work. Do not invent a universal team and do not treat generated role files as proof of expertise.

Eligibility gate

Confirm that the user already has Codex, Claude Code, or CodeBuddy Code running and can open a local folder. If not, stop: application installation and onboarding are outside this product.

Obsidian is optional. The generated files are ordinary Markdown and JSON; Obsidian is only a viewing and linking interface.

Workflow

1. Read interview.md and conduct the interview in short batches.
2. Classify each answer as confirmed user fact, requested behavior, or assistant recommendation.
3. Propose the smallest role set. Always include one coordinator. Add a specialist only when recurring work and a distinct quality test justify it.
4. Confirm the target platform: codex, claude-code, codebuddy-code, or universal. Do not label Tencent WorkBuddy as a native target.
5. Show the proposed roles, permissions, sensitive-data boundaries, sources of truth, platform adapters, and files before writing.
6. Save the approved answers as a JSON profile matching profile-schema.md.
7. From the Brain Kit root, run python skills/setup-brain/scripts/brainkit.py preflight --platform <platform> --target <folder>.
8. Run python skills/setup-brain/scripts/brainkit.py plan --platform <platform> --profile <profile.json> --target <folder>. Present the plan and collision report.
9. Only after explicit approval, run python skills/setup-brain/scripts/brainkit.py apply --platform <platform> --profile <profile.json> --target <folder>.
10. Run python skills/setup-brain/scripts/brainkit.py validate --target <folder> and report every warning. A successful install is not proof that any role is trained.
11. If the user requests reversal, run python skills/setup-brain/scripts/brainkit.py rollback --target <folder>. A partial rollback preserves changed files and the manifest for later review.

Non-negotiable boundaries

• Never overwrite an existing file, including brain/install-manifest.json. A collision aborts apply.
• Reject duplicate/reserved role IDs and unsafe manifest paths before any mutation.
• Never copy the developer's client projects, personal data, role names, memories, credentials, or absolute paths.
• Never create more roles merely to make the system look professional.
• Keep Skill, Knowledge, Evidence, Decision, and Project Fact separate.
• Multi-agent agreement is not evidence.
• External messages, publication, payments, deployments, and destructive actions require task-specific approval.
• Generated specialists begin at untrained; they advance only through sources, tests, and project validation.
• Dynamic facts such as prices, policies, account state, and software versions require runtime verification.
• Keep shared rules in AGENTS.md; platform adapters import or route to it rather than duplicate project facts.

Output contract

The generated workspace must let a new task answer, without reading every file:

• Who is coordinating?
• Which role owns this task and who challenges it?
• What may be decided autonomously?
• Where do project facts, evidence, knowledge, and decisions live?
• What is untrained or unverified?
• Which platform adapter is active?
• What changed during installation, and how can it be rolled back?

Use acceptance-tests.md for the final quality gate.