Agent-led elicitation¶

Definition¶

Agent-led elicitation reverses the usual direction of instruction: instead of the person specifying a task they do not yet understand, the model interrogates the person — mapping the traps and questions they did not know to ask, or interviewing them one question at a time — so that the requirements are produced by the exchange rather than assumed to exist beforehand.

Explanation¶

The mechanism rests on an asymmetry of coverage. A model has seen far more of the average domain than the person asking about it and can read a codebase or search the literature in seconds, so it can enumerate the relevant question set faster than a newcomer can discover it by trying things. Two shapes recur. The blind-spot pass is a one-shot request to name what you do not know you do not know about a specific area — historical work already done, standard pitfalls, what good even looks like — and it is most useful precisely where you have the least standing to write a good prompt. The interview is iterative and needs a prioritisation rule to be worth the turns; asking for one question at a time, ranked by which answer would most change the architecture, filters the exchange down to decisions that are expensive to revisit later. Both depend on the human declaring their starting point — what they have tried, what they already know, how much of the codebase is familiar — because the productive questions for a novice and an expert in the same task barely overlap. The cost profile is what makes this a discipline rather than a nicety: the questions are cheap now and the same discoveries arrive expensively later, as rework, if nobody asks them.

Key Properties¶

The model asks and the human answers, inverting who holds the specification
Two shapes: a one-shot blind-spot pass, and an iterative one-question-at-a-time interview
Interviews need a prioritisation rule — questions whose answers would change the architecture come first
Requires the human to state their own expertise and starting point, or the questions land at the wrong level
Works because model breadth exceeds the newcomer's, not because the model knows the specific situation

Relationships¶

Unknowns inventory in prompting — is the framing this technique serves, since it targets exactly the quadrant a person cannot query on their own
Disposable artifacts as specification — complements it by addressing criteria that are tacit rather than absent — questions surface what you never considered, artifacts surface what you could never articulate
Rare context — agent-led elicitation is a concrete technique for surfacing exactly the tacit knowledge rare context names — interrogating the person for local vocabulary and conventions they would not think to volunteer unprompted.

Applications¶

Opening work in an unfamiliar module, domain, or craft — a new auth subsystem, a discipline like colour grading — and pre-flighting a spec by having the model interview you about ambiguities until the architecture-relevant answers are settled.

Sources¶

https://claude.com/blog/a-field-guide-to-claude-fable-finding-your-unknowns