status: active
timestamp: 2026-07-02
tags: [agent, prompting, context, interview, uncertainty]
Context interview — agent asks user first when uncertain
When unsure what context is needed for a task, agent asks the user targeted questions BEFORE attempting. 'Ask me any further questions you need to achieve the best result.'
Context interview
The rule
When the agent is uncertain what context is needed to produce a good result, ASK THE USER first. Don’t guess. Don’t proceed with vague framing hoping to iterate.
The pattern
Agent detects uncertainty → sends 2-4 targeted questions → user answers → agent proceeds with grounded context.
Example prompt to the user when the AGENT is asking:
“Before I write this, I need to understand a few things. Answer any that are relevant:
- Who is the audience?
- What’s the desired tone?
- Are there constraints (length, format, forbidden words)?
- Are there examples I should model after?”
When to fire
- Task is broad or ambiguous (“write a blog post” without topic angle)
- Task involves taste (design, prose, code style beyond formatter rules)
- Task references undocumented context (“the usual way”, “like last time”)
- Task has 3+ valid interpretations
When NOT to fire
- Task is mechanical (rename files, apply lint rule, run test)
- Task has clear spec in
knowledge/— use spec, don’t interview - Task is a follow-up in same session where context was established
- STT-friendly ambiguity (per
communication-stt-friendly) — pick most likely interpretation + state assumption inline, don’t interview
Not the same as grill-me
Grill-me = user invokes decision-tree interrogator to lock a choice with multiple locked-in options. Context interview = agent asks small factual questions to fill in gaps before executing.
Grill-me is user-invoked, uses MCQ, ends in a knowledge write. Context interview is agent-invoked, uses open questions, ends in a well-scoped task attempt.
Format
- Numbered questions — user can answer subset by number
- ≤4 questions per interview — SDK MCQ cap + STT-friendly
- Explicit “answer any that are relevant” — user can skip
Anti-patterns
- ❌ “What do you want?” — too broad, user has to redo the framing
- ❌ 10-question interview — pick 4 highest-leverage
- ❌ Interview after already producing wrong output — should have interviewed first
- ❌ Interview mid-task when agent should have asked at start
- ❌ Skipping interview + producing wrong output + saying “let me know if that’s not right” — that’s punting the interview onto the user
Cross-refs
karpathy-guidelines§Think Before Coding — state assumptions; interview is one way to gather themedit-mode-prefs— “no report/summary .md files as deliverable” applies; return context questions inlinecommunication-stt-friendly— the interview may itself be STT-friendly; ≤4 short questions- Source: Nate Herk YouTube 2026-07-02 — “just ask Claude to ask me any further questions it needs to achieve the best result”