Turning the interview around: using LLMs to extract requirements
Large language models need context before they can take on complex jobs. A design task might require user-facing expectations, implementation constraints, references to external systems — often several pages of markdown before the model can start. The standard route is to have a human draft all that context. An alternative is to delegate the drafting to an LLM that pulls the information out of a human by asking questions.
The pattern is straightforward to put in place. You prompt the LLM to interrogate you, letting it ask whatever questions it needs to write the contextual brief. You answer, point it toward further sources where relevant, and when the interview is over the model produces the context report for another session — possibly on a different model — to do the actual work. The approach was described in detail by Harper Reed, whose workflow includes the notable demand that the LLM ask exactly one question at a time. In practice, it takes frequent reminders to keep the model on that rule.
Document validation through conversation
A second flavor of interrogatory LLM starts with a document rather than an empty page. Hand it something like a software specification, and ask it to interview a human expert to verify whether the document reflects reality. This can replace the usual review process in which an expert is asked to read the document and judge its correctness. Reading a dense or poorly written spec is a chore many people avoid, and a dialogue with an LLM can surface issues that a passive read would miss.
The two variants compose cleanly: use one interrogatory LLM to draft a document, then run additional interrogations with other experts to check it.
This technique goes beyond LLM-focused tasks. Because not everyone thinks best by writing, the interrogatory format offers a channel for people who find writing difficult to get their knowledge out of their heads and into reusable form. The output will carry a recognizable AI voice, which may offend some readers — but an imperfectly phrased document is still preferable to losing the information entirely to rushed or skipped writing.
