Less Expertise, More Coverage

The intuition we borrow from people

If you have a hard analytical question, you call the most specialized person you can find. A domain expert sees more, cuts faster, wastes less time on dead ends. That instinct is correct for people, and it is the first thing most of us reach for when we write a prompt. “You are a senior tax attorney. You are a clinical pharmacologist. You are a fraud analyst with twenty years on the desk.” The framing feels like it should sharpen the model.

It does sharpen the tone. The output reads more confident and more fluent. The question I kept coming back to is whether it actually sharpens the analysis, and whether confidence and coverage move in the same direction. They do not always.

What the expert framing does to coverage

This is the claim in my SSRN working paper, Less Expertise, More Coverage: The Counterintuitive Effects of Prompting LLMs for Analytical Tasks. When you cast a model as a narrow domain expert, you do not just change its vocabulary. You change what it treats as in-scope. A “senior tax attorney” answers like one: it follows the well-worn paths an expert would, and it skips the angles an expert has learned to ignore. For many analytical tasks the goal is not the single best take. The goal is coverage of the possibility space, the long tail of considerations that a real answer has to account for.

The expert prompt optimizes for the confident center. The task often lives in the edges it decided were beneath an expert.

A broader, less-constrained framing tends to keep more of that tail in play. It is messier. It surfaces considerations that a specialist would dismiss, and some of those are noise. But on a task where missing a branch is the expensive failure, the broader prompt covers more of the real answer than the expert one does. I want to be careful here: the paper measures this effect rather than asserting it, and the size of the gap depends heavily on the task. I am not claiming the expert framing is always worse. I am claiming the assumption that it is always better is wrong.

Breadth and depth are different jobs

The cleanest way I have to hold this is to separate two questions. How deep is the model on any single line of reasoning, and how much of the space does it cover at all? Expert framing trades the second for the first. You get a deeper, more polished take on a narrower slice. For a focused execution task that is the right trade. For first-pass analytical work, where you are still mapping what the question even contains, it is the wrong one, because the failure mode is not a shallow answer. It is a missing branch you never knew to ask about.

This is also where the danger hides. The expert prompt produces the version that sounds the most authoritative while quietly covering the least. A reader has no easy signal that whole regions of the problem went unexamined. Fluency and coverage have come apart, and only one of them is visible.

Why this shows up in research tooling

I work on a research platform at Articos, and this finding lands directly on how I design prompts and protocols there. Early-stage hypothesis generation is a coverage problem before it is a depth problem. When the platform is mapping what could be true about a set of users, I do not want a confident expert voice that prunes the space down to the obvious. I want a wide first pass that names the branches, and structure around it to test them. The expert costume comes later, when a specific branch is worth going deep on.

Two mechanisms I have written about elsewhere matter here. The first is calibrated confidence: a finding carries a confidence signal a reader can weigh, so the authoritative tone of an expert framing can never substitute for evidence. The second is the audit chain behind auditable AI research, which makes it possible to see which branches a study actually explored, not just how the conclusion reads. Together they stop the expert framing from doing its quiet damage: you can no longer mistake a narrow, confident answer for a covered one.

Falsifiable, not clever

I am wary of prompting advice that is a pile of tricks. What makes this useful is that it is testable. Pick an analytical task with a known answer space, run the expert framing against a broad one, and measure coverage rather than how good either reads. If the expert prompt covers more, I am wrong for that task, and I would rather know. That is the same standard I hold the research platform to. The point of the paper is not a slogan about expertise. It is a measurable effect you can check against your own work, and a reason to stop assuming the expert prompt is free.

The full argument and the measurements are in my SSRN working paper, Less Expertise, More Coverage: The Counterintuitive Effects of Prompting LLMs for Analytical Tasks. I’ll link the live paper here once it is indexed. The architecture this connects to is in Grounded Simulation, which is where coverage and auditability get built into the method rather than bolted onto a prompt.