---
name: auto-intake
description: Auto Office 3.0 reference spoke, not loaded by 3.1 runs (they use the office CLI). Internal Auto Office v3 intake spoke. Use when the planner runs the interactive requirements interview before freezing intent — covering the twelve-item floor in one or two batched question rounds (or a structured-question tool, per harness) and deriving the five frozen fields from the answers. This spoke is planner-owned and requires talking to the user directly. Do not use to ask about gear, which is declared, not interviewed.
---

# Auto Intake

> **Auto Office 3.1:** this is 3.0 reference material. A 3.1 run is driven by the `office` CLI and runtime-delivered
> role briefs; do not run the `office_runtime.py` helpers below for it. Follow `office status` and its `next:` line.

The planner conducts this interview directly with the user, after repository reconnaissance and using the orchestrator's provisional intent only as a starting hypothesis it may revise. Use the harness's native question tool when it has one (the per-harness list is under "Asking the user" in the root `SKILL.md`; Claude Code: `AskUserQuestion`); batched plain text otherwise — either form satisfies intent coverage as long as all twelve items are covered before freezing.

## The twelve-item floor

Cover all twelve on every run, delivered as one or two batched question rounds, never a back-and-forth conversation:

1. **Outcome** — what is true when this is done, in the user's words.
2. **Done-criteria** — exact commands, reads or observations that prove it.
3. **Blast radius** — repos, environments, live systems; production named or excluded explicitly.
4. **Irreversible steps** — each with preconditions written out.
5. **Waves** — which done-criteria can proceed simultaneously vs. must serialise.
6. **Interfaces** — signatures, schemas, routes, file boundaries parallel tasks must agree on.
7. **Constraints** — stack, conventions, domain skills, untouchables.
8. **Speed vs correctness** — which one is being bought.
9. **Executor count** — one repo or several, one slice or several.
10. **User-owned decisions** — anything that would otherwise be guessed.
11. **Rollback target** — what "undo this" concretely means.
12. **Prior art** — existing work, branches or PRs this must not duplicate or contradict.

## Freezing the five fields

`goal`, `done_criteria`, `blast_radius`, `named_actions`, and `non_goals` are **derived** from the twelve answers, not asked directly. Every frozen value traces back to an answer above; one that doesn't is the defect this spoke exists to prevent. Record the freeze with `python3 scripts/office_runtime.py freeze-intent --state-dir <dir> --intent <json>`; it is the only intake → planned transition, and `approve-plan` refuses any other phase.

## Irreversible steps become named_actions

Item 4 becomes a `named_actions:` entry with its preconditions written out exactly (`{"action": ..., "preconditions": [...]}`; `freeze-intent` rejects anything less) — that's the receipt that lets the loop perform it later without stopping. Preconditions you can't yet write exactly mean the interview isn't finished; keep asking.

External sends always stop the loop, regardless of preconditions — never a `named_actions` entry.

## Exit test

The receipt is frozen intent a stranger could build the right thing from: a routed executor in a separate process, with no access to this conversation and no chance to follow up. Anything they'd have to guess means the interview isn't done.

## Gear is not a thirteenth question

Gear is declared in the kickoff block, decided by the fit test from blast radius, reversibility and size class. It ships from there only — this interview never asks for it.