Interview Me
The user wants to answer your questions in the agent's interactive question UI, not in chat. You show each question as a picker. The user moves between questions or answers with the arrow keys and picks an answer, or types their own in the free-text field.
The one rule: never write a question in chat text. No lists of questions, no "(a)/(b)/(c)" options in a message, no "Let me know which you prefer". If you have a question for the user, it goes through the question tool.
Which tool
| Agent | Tool | Free-text option | Checklist (several answers) |
|---|---|---|---|
| Claude Code | AskUserQuestion | added automatically ("Other" / "Type something") | multiSelect: true |
| Codex | request_user_input | added automatically ("Other") | not supported, see below |
Don't add your own "Other" or "Something else" option. The tool already adds one, and a second one is clutter.
If neither tool is available, see When the tool is missing.
What to ask about
- A topic given after the command (e.g.
/interview-me about the database schema): interview the user on that topic. Work out what you need to know first. Read the relevant code or docs so your options are concrete and grounded, not generic. - No arguments, and you have open questions (the usual case: you just asked something inline, or you are about to make a call that is the user's to make): put those questions to the user through the tool. Don't add new ones.
- No arguments, nothing pending: use the tool to ask what they want to be interviewed about. Offer 2–4 likely topics taken from the conversation so far.
- Not invoked, but you are about to ask the user something inline: ask it through the tool instead. This only changes how you ask. Don't open an interview just because a decision is coming up, and don't add questions you wouldn't have asked.
Don't announce the mode or explain the rules. Go straight to the first tool call.
Writing good questions
- One decision per question. Split compound questions ("which DB and should it be hosted?") into separate ones.
- Self-contained. The user may see only the question dialog, not your earlier text. Each question should name what it is about and, briefly, why it matters ("The importer keeps duplicates. How should it handle them?").
- Short header, max ~12 characters (e.g.
Auth,Storage,Scope). - 2–4 real options (Codex: 2–3), each with a label of 1–5 words and a one-line description of what choosing it means or costs.
- Options must be mutually exclusive for single choice. If they aren't, it's a checklist question.
- Recommend when you can. Put your recommended option first and end its label with
(Recommended). Only do that when you actually have grounds for it. - Use previews for visual comparisons (Claude Code): when options are code
snippets, layouts, config variants or diagrams, add a
previewto each option so the user can compare them side by side. Previews only work for single choice. - Don't ask what you can find out yourself. Facts in the code, the docs or the conversation aren't questions. Ask about intent, preference, priorities and trade-offs.
Single choice vs. checklist
- Use single choice when exactly one answer can apply (pick a library, pick an approach).
- Use a checklist (
multiSelect: truein Claude Code) when several can apply together (which features to include, which platforms to support, which files to touch). Phrase it that way: "Which of these should the export include?" - Codex has no checklist. Use single choice with the most useful combinations as
options (e.g.
CSV only,CSV + JSON,All formats), and say in the question that the user can type a custom combination in the free-text field.
Running the interview
- Batch related questions: up to 4 per call in Claude Code (3 in Codex). Put questions that don't depend on each other in the same call. A question whose options depend on an earlier answer goes in the next round.
- Read the answers literally. Free-text answers are real answers. If one is unclear or opens a new branch, ask a follow-up through the tool.
- Keep going in rounds until you have what you need. Don't pad with questions you don't need, and stop as soon as the open decisions are settled.
- Never ask "is that all?" or "shall I proceed?" through the tool. When the
interview is done, give a short summary in chat, one line per decision
(
Storage: SQLite (you: "keep it single-file")), then carry on with the task. In plan mode, finish the interview before presenting the plan. - If the user declines or dismisses a question, don't ask it again. Go with your best judgment and say in the summary what you assumed.
The mode ends with the interview. After the summary you go back to normal behaviour, until the user asks for an interview again (or you are about to ask questions inline, see case 4 above).
When the tool is missing
Check your tool list, don't guess.
Codex, Default mode: request_user_input is normally only available in Plan
mode. Tell the user in one line and give them both ways to fix it:
- Switch this session to Plan mode with
/plan, then run$interview-meagain, or - Enable the tool in Default mode (feature flag, still marked "under development"):
or start Codex withtoml# ~/.codex/config.toml [features] default_mode_request_user_input = truecodex --enable default_mode_request_user_input.
codex exec / non-interactive runs: the tool doesn't work there at all. Say so.
Any other agent without a question tool: say that interactive questions aren't available here.
In all of these cases, don't silently fall back to a wall of inline questions. After the one-line notice, ask only the first question in plain text with lettered options and an "or type your own" line, wait for the answer, then ask the next one.
