Ambition is the bottleneck, not the model

The ceiling is in the request

Most people use a strong model on small, one-off tasks, and ask for what they already know how to ask for. The model does exactly that. So the quality of the answer is capped by the ambition of the question, not by what the model can do.

That is a shift. A few years ago the limit sat in the model, and prompting was a way of working around it. Today the limit more often sits in the person typing. If you ask for a summary, you get a summary, even when the model could have rebuilt the analysis behind it.

Four knobs

I think of an ambitious request as four settings you can turn.

Speed. How much of the work you hand over at once, and how fast you expect it back. A model that can do an afternoon of work in minutes is wasted on a one-line question.

Altitude. Aim above the literal ask. The request names a task, but the goal behind it is usually bigger, and the answer should serve the goal.

Depth. Start from raw material, not from someone else's digest. A model working from a summary inherits every choice the summarizer made. A model working from transactions, transcripts or source documents makes its own.

Range. Name the methods and tools an expert in the field would reach for, not just the ones you happen to know.

The trouble is that turning these knobs requires knowing what an expert would do. Often you don't. That is the gap the next part closes.

Ask the model what you didn't know to ask

Append this to any non-trivial request:

Before you do the work, answer four questions. (1) What would a top professional in this field do that I'm not asking for? (2) What is the rawest material they would start from, instead of already-digested material? (3) Which specific methods and tools would that expert use? (4) What, other than you or me, would show the result is right? Then rewrite my prompt to ask for exactly that, show it to me, and run it after I approve.

The model knows far more about professional practice than any single request draws out. This prompt makes it spend that knowledge on the framing of the job, before spending it on the job.

The fourth question does the real work

Questions one to three raise the ambition. Question four changes the kind of answer you get.

Without it, the only judges of the result are the model and you. The model writes fluent text whether it is right or wrong, and you are reading it quickly. "Sounds good" is the whole test.

An external check replaces that with something that can fail. A reconciliation against the ledger, a test suite, a primary source, a second dataset, a calculation done a different way. When the model has to name that check up front, the work gets built to pass it, and you get a way to find out if it didn't.

Here is the shape of it. Asked to analyze a sales export, a model summarizes it. Asked what an analyst would check, it rebuilds revenue from the raw transactions and reconciles the total against the ledger.

Where it breaks

It is not for everything. A simple task, such as rewording an email or converting a date format, gains nothing from four questions and a rewritten prompt. You only add cost.

And it does cost. More thinking before the work means more tokens and more time. For a task you run once a day, that adds up.

The larger risk is scope. Asked what a top professional would do, a model can describe a week of work for a task that needed an hour. It may propose tools you don't have or checks you can't run. That is why the prompt ends with an approval step. The rewrite is a proposal. You read it, cut what you don't need, and only then let it run. Skipping that step hands your judgment of the scope to the model.

What to do this week

Pick one task you would normally ask in a single line, something that matters enough to be checked. Paste the prompt above under it. Read the rewrite before approving, and look hardest at question four. If the check it names is one you can actually run, you have changed what the answer is worth.

To make it permanent for a given area of work, put the prompt in the instructions of a Claude Project, so it applies to everything you do there.