Technology RadarTechnology Radar

Elicitation-First Prompting

workflow
Assess

Elicitation-first prompting is the practice of using an AI model as an interviewer before using it as a writer. Instead of asking AI to generate content directly, you ask AI to ask you questions — then answer freely, and use the transcript as the raw material for writing or specification. The insight: your best thinking emerges under interrogation, not from a blank page.

The Core Technique

Articulated by Luke Bechtel in "What Only You Can Say" (2025):

  1. Prime the AI as an interviewer — e.g., "Act as a skilled editor preparing to help me write about X. Ask me questions one at a time, probe my answers, push back when something is vague. Don't write anything yet — just interview me."
  2. Answer freely — ramble, contradict yourself, backtrack, follow tangents. The goal is to surface what you actually think, not to produce polished output.
  3. Use the transcript as raw material — the interview log becomes a record of your thinking. Now ask AI to draft from that, not from nothing.

The phrase "what only you can say" names the thing at stake: the specific experiences, unexpected formulations, and genuine contradictions that emerge mid-answer. AI-generated prose is competent but generic; it cannot produce the moment where you say "actually, the real point is—" — that inflection is irreplaceable, and it only surfaces under questioning.

Why It Matters for Agentic Engineering

This pattern reframes the human's role in human-AI collaboration from approver (reviewing what AI produced) to primary knowledge source (generating the irreplaceable material that AI then structures).

In software development contexts, elicitation-first maps onto:

  • Pre-spec interviews — before writing a technical spec, have AI interview you about requirements, edge cases, and constraints. The interview surfaces assumptions you didn't know you had.
  • Pre-mortem dialogues — "interview me about what could go wrong with this design." More effective than asking AI to generate failure modes it has no context for.
  • Architecture decision records — the interview transcript often contains better rationale than a blank ADR template prompts you to write.
  • Commit message and PR description drafting — "Ask me questions about what I changed and why before writing the description."

Why Assess, Not Trial

The technique is promising and the underlying insight is sound — but:

  • It requires discipline to use consistently; the path of least resistance is still to ask AI to draft directly
  • The quality gain depends heavily on how much domain knowledge lives in the practitioner's head, not in available documentation
  • Best documented for writing and communication contexts; agentic engineering applications are largely extrapolated, not yet widely reported with outcomes

It warrants experimentation, particularly in design review and specification workflows where AI-generated first drafts are most prone to generic output.

Key Quote

"Your best ideas come from conversations you weren't planning to have." — Luke Bechtel

See Also