Interrogatory LLM
What it is
A pattern for using LLMs not as answer generators but as interrogators—systems that ask clarifying questions to surface hidden assumptions, ambiguities, and tacit knowledge in a user's request.
Coined by Martin Fowler, the term draws on linguistic theory: interrogative mood (asking questions) vs indicative mood (stating facts). Most LLM usage is indicative—"here is my question, give me an answer." Interrogative usage flips the script: "here is my situation, ask me questions until we both understand what I actually need."
Why it matters
The most expensive failures in AI-assisted work come from misunderstood requirements, not from bad generation. A small ambiguity at the start of a multi-hour agent session can compound into massive wasted effort. The interrogatory pattern catches these ambiguities early.
Fowler's "missing half"
Fowler argues that if indicative LLM use is currently a 10, interrogatory use is a 1. The field is dramatically underinvested in AI systems that help humans think by asking better questions.
Key mechanics
- Requirement extraction via dialogue: Instead of the user providing a complete spec upfront, the LLM interviews the user to build a shared understanding.
- Assumption surfacing: The LLM explicitly asks about edge cases, constraints, and preferences the user hasn't stated.
- Ambiguity flagging: When the request contains multiple valid interpretations, the LLM presents them and asks which the user intends.
Related concepts
- Judgment Engineering — both address the need to make tacit knowledge explicit
- Interactive Requirement Extraction — Anthropic engineer Arnaud Doko's similar practice
- 百问计划 — the "100 questions" approach to requirement clarification
Open questions
- Can interrogatory patterns be standardized, or do they need to be domain-specific?
- Does the benefit justify the extra interaction rounds, or does it slow down experienced users?