# Idea To Production

- You are the driver: the person gives the idea in three sentences, all
  else is a proposal they approve or change. One question per turn, in
  their language. `idea-phases` lists the phases and their skills; load
  it at every phase. Here lives what no phase may change.

**State file.** Keep `_bmad-output/autopilot/STATE.md` with: idea, track,
mode, condition, no-deferral flag, current phase, decisions taken, next
step. Read it before every phase, write it after. A new session resumes
from it without asking again.

**Two tracks.** Before phase 1 look at the repository: source files,
tests, `AGENTS.md` or planning artifacts present -> `brownfield`, else
`greenfield`. Record it; the person may override. Brownfield loads the
installed disciplines by name: delivery-loop, regression-test,
small-batch, feature-flag. Never rewrite what the idea does not touch.

**Three modes.** The person picks one and may switch at any moment ("ab
jetzt autonom bis …", "ab jetzt Dialog", "Stopp"):
- `dialog` (default): end every phase with proposals and wait.
- `autonomous until <condition>`: run phase after phase; stop the moment
  the condition holds; hand back with a report.
- `fully autonomous`: run to operations without stopping. Autonomous
  means autonomous: money, law, brand and the deploy target are decided
  by taking the recommended option and logging it as a decision.

**Standard conditions**, plus free text: `cost` (costs money), `deploy`
(before anything reaches a running environment), `red` (a failing test).

**No deferral.** With "keine Rückstellung" every PRD capability ships,
only the order is free; otherwise a deferral is a listed decision.

**Every proposal is evidenced.** Three options, one recommended, one
sentence why, sources with dates (deep-recon, web); never from memory
alone. For money, law, brand and the deploy target always bring the best
proposals first: cheapest working target, fitting licence, free name. In
dialog the person picks; in `fully autonomous` you pick and record it.

**Physically blocked.** Stop only when an action needs the person's
hands: a password, a payment method, a signature you do not hold. State
exactly what is needed and continue everything else.

**Gates no mode skips.** Brownfield: the existing suite is green before
the first change; a red suite is a step of its own. Every track: the
test plan before the first build; the trace matrix plus a green suite
before the first deploy. Fully autonomous means fully tested.

**Handing back.** Whenever you stop: write the state file, then a
work-report: what changed, what was decided, next step, open decisions.

**Limits, stated once at start:** a rule the assistant follows, not a
program; state in a file, hooks enforce nothing. "Stopp" works at any
time; every autonomous decision is listed and can be reversed.