interview-me

Interview Me Skill

The most expensive failure mode in AI-assisted work isn't bad output — it's excellent output to the wrong brief. This skill inverts the flow: before producing anything, interview the requester like a senior consultant would, one question at a time, until the brief can survive contact with the deliverable.

What This Skill Produces

• A validated brief: goal, audience, constraints, success criteria, non-goals — confirmed by the requester
• Then the actual deliverable, built against that brief
• A visible assumption ledger for anything the interview couldn't settle

When to Trigger (and when not)

Interview when: the request is one sentence for a multi-hour deliverable · the audience or purpose is unstated · two readings of the request lead to different artifacts · the stakes are high (board, customer-facing, irreversible). Skip the interview when the request is already specific, the pattern is established from earlier in the conversation, or the cost of a wrong draft is lower than the cost of five questions — say "I have enough to start" and start.

Interview Method

1. One question at a time. A wall of seven questions gets skimmed answers to all and real answers to none. Ask, absorb, let the answer shape the next question.
2. Sequence by decision-weight. The order that converges fastest:

• The moment of use — "who reads/uses this, and what are they doing in that moment?" (settles more downstream decisions than any other question)
• The definition of success — "what happens if this works? what would make you send it back?"
• The constraints that bind — length, tone, format, deadline, politics ("anything this must NOT say?")
• The prior art — "has something like this been tried/shown before? what happened?"
• The non-goals — "what's adjacent that we're deliberately not doing?"

1. Interrogate the difference, not the topic. Weak: "tell me more about the dashboard." Strong: "if this dashboard existed today, what decision would someone make differently this week?"
2. Offer forks, not open fields, when the requester is fuzzy. "Is this closer to (a) a live monitor the team glances at, or (b) a monthly readout for your boss?" — concrete options unstick vague askers far faster than "what do you envision?"
3. Know when to stop. 3-6 questions settles most briefs. Stop when a new answer wouldn't change what you'd build. Then play the brief back in ≤5 lines and get an explicit "yes, build that."
4. Ledger what's still open. Unresolved items become labelled assumptions in the deliverable, never silent guesses.

Output Format

During: one question per turn, with a one-line reason when it isn't obvious ("asking because it changes the format entirely").

The brief playback:

> Building: [artifact] for [audience in their moment] so that [the decision/outcome].

> Success: … · Constraints: … · Not doing: …

> Assumed (unconfirmed): …

> Confirm and I'll build it.

Quality Checks

• [ ] Questions were asked one at a time, each shaped by the previous answer
• [ ] The moment-of-use and success-definition questions were asked (or their answers were already known)
• [ ] The brief was played back and explicitly confirmed before production began
• [ ] Every unresolved item appears in the assumption ledger, labelled
• [ ] The interview stopped when answers stopped changing the build — no ritual questioning

Anti-Patterns

• [ ] Do not fire a questionnaire — seven questions at once produces skim-answers and resentment
• [ ] Do not interview when the brief is already clear — process applied without judgment is friction
• [ ] Do not ask questions whose answers wouldn't change the deliverable — every question spends the requester's patience
• [ ] Do not start building mid-interview "to save time" — half-brief work anchors the requester to the wrong draft
• [ ] Do not skip the playback — the interview's value is captured only when the requester says "yes, that"