Mine Your Work History Into an Agent Team

Premise

Your machine already contains the evidence for which agents you need: repeated tasks, reused prompts, ignored tools, and recurring corrections. Convert that history into files, not another aspirational org chart.

Setup

1. Create an isolated team-build/ output directory.
2. Approve which configuration and project folders may be inspected.
3. Exclude secrets, credentials, browser data, and unrelated personal files.
4. Require a read-only end-to-end verification before any external action.

Boundaries

• Secrets are never opened or printed; only excluded filenames may be noted.
• Sending, paying, publishing, deploying, deleting, and connecting accounts require approval.

Work in order: (1) inventory approved agent configs, project rules, repeated prompts, recurring tasks, corrections, successful outputs, and abandoned attempts; (2) classify findings into skills, integrations, voice rules, possible roles, and junk; (3) design minimal roles with one owner, prohibitions, and approval gates; (4) write role and skill files; (5) verify one read-only task from routing through saved output; (6) produce HANDBOOK.md.

Record my future corrections in the relevant rule or skill file. Evidence of completion must be a path, diff, test, or verified response.

Never inspect secret contents. Stop before any send, payment, post, deployment, deletion, or external connection.

Build the smallest useful agent team from evidence on this machine. Write every artifact beneath `team-build/`; nothing important may exist only in chat.