Interviewing the Operator¶
Source: hooks/ways/meta/choices/interview/interview.md
Frontmatter
| Field | Value |
|---|---|
description |
interviewing the operator on a series of open directional decisions, such as a draft record's open questions or a design's forks, each question carrying a brief, the case for and against, and a lean defended from checkable evidence, and ending when the operator trusts the agent to carry on alone |
vocabulary |
interview operator directional open questions lean evidence brief for against judge micro-decisions microdecisions framing round series satiation consequence defend grounded accept leans hold research unknowns handoff get it done carry on alone sufficient trust stop conditions |
pattern |
\bopen questions?\b|\binterview (me|the operator)\b|\bmicro-?decisions?\b |
refire |
0.15 |
scope |
agent, subagent |
When several calls above the waterline (choices) are open at once, put them to the operator as a series of questions, each with the evidence to judge it by.
Before the questions¶
Triage first. Settle the calls below the waterline yourself or with subagents, and list each with your lean in a line the operator can overrule. If you cannot write a real case against your lean, the call is yours. A call made under high doubt (core.md) goes to the operator even below the waterline.
Then frame the set in the reply as a markdown table with numbered rows, so an answer can name rows: each open decision, what it changes in the finished work, and your order. Keep artifact pages for things the operator looks at or uses.
Each question¶
Each question has three parts:
- A brief in plain words: what the setting or rule does and what a user would see under each option, with an everyday example. A newcomer to the domain can follow it. Leave out mode numbers, protocol names and spec terms, or define each in a few words.
- The case for the leading option, and the case against it.
- Your lean, listed first and marked (Recommended), with evidence the operator can check: a file and line, a record, a measurement, a spec, a precedent. A lean with no evidence is a defect: find some, or say the call turns only on what the operator wants.
Keep a question to about a hundred words, with detail moved to the framing or the record. choices has the other presentation rules.
Name the unknowns your evidence leaves open. Offer a research pass before asking where a call needs knowledge the operator may lack, and a "hold, I'll research first" option where they may want to look before committing, as between libraries. A hold pauses only the work that depends on the call.
How to ask: by whether the operator is present¶
The question tool (AskUserQuestion) puts each question as options with context, your lean first, and room for the operator to answer in their own words. That shape draws a clearer answer than prose does, and a clearer answer quotes cleanly in a record (adr/consider). The tool holds your turn open until the operator answers, so start every piece of work that does not depend on the answers before you ask. Treat the operator as present when they have written in this session and no /goal run, schedule or grant to carry on applies.
- The operator is present. Ask through the question tool, one round at a time. Where the session has no question tool, as in a headless run, ask in your reply in the layout of step 3 and end the turn.
- The operator is away and listens on attend. Post the questions to them as a direct message, laid out so "1 yes, 2 hold" is a full answer. Their answers arrive as messages while the work continues.
- No one is present. Put the questions in your reply as a numbered table with each question's brief, the case for and against, and the lean. Keep working on what does not depend on the answers, and act on a stated lean only where the step is easy to reverse. Under a grant, decide by your leans and log each call (Ending in a handoff).
An answer that overturns a lean you acted on is applied as a revision, and you say what changed.
Rounds¶
- Ask one to four questions per round, the call that most changes the work first.
- Offer a way to accept the remaining leans in one answer.
- Word each question from its own decision. A stock sentence repeated across a batch stops being read.
- Re-triage after each round, since an answer can settle a later question or raise a new one.
After the answers¶
Record each answer where the work keeps it: adr consider N --said "..." --via "..." for a decision record, quoted as adr/consider says, or a comment on the PR or issue otherwise. Then act, and say which leans were taken, which were overruled, and which calls are on hold.
Ending in a handoff¶
When more calls above the waterline will come, the last round may ask whether to keep asking or carry on alone. First check your read of the operator's intent in up to four questions:
- The goal and what done means, in two or three lines.
- The leans you will take on the calls ahead.
- What sends you back to ask. Offer as defaults a one-way step, an action others see outside the working branch, spending money or quota, and anything outside the stated scope. Rising doubt (core.md) stays on the list under every grant.
- How you will report: a log, in the PR or the record, of each call above the waterline you made alone, with its lean and evidence, for review at the end.
Then ask for the mode: keep asking, or carry on, optionally bounded by scope or time. Mark neither option as recommended. A "no" or no answer means keep asking (trust/autonomy).
A grant covers this one piece of work until it is done or a stop condition is reached. Under it, decide the remaining calls by your stated leans, logging each. Only checks you have seen fail on a planted defect (core.md) count toward low doubt. In a decision record, cite the grant as the operator basis (adr/consider). For work that runs unattended across turns, write the grant as a /goal condition (goal-author).
See Also¶
- choices(meta) — parent: the waterline, and presenting a single choice
- adr/consider(documentation) — the operator's check of a drafted decision record
- trust/autonomy(meta) — the operator grants autonomy; the agent never argues for it
- goals(meta) — a grant that runs unattended as a
/goalcondition - think(meta) — framing an unknown your tools cannot resolve