110

One level up

#
  • heuristic
  • opinionated
  • no code

When two or three requests in a row have not changed your understanding of the problem, stop refining and step up: from the request to the task, from the task to the problem.

A stalled dialogue looks productive: the model answers at length, agrees with every refinement and offers a next step. The cause is nearly always the same — the task is being solved at the level it was posed, and it was posed above or below where the problem lies. The paragraph is not bad; the thought behind it is. Delivery ads are not the reason orders fall. The model will rewrite the paragraph forever, because that is what it was asked. Rising agreement without a shift in understanding is the signal.

The costliest move here: admitting the framing you came with was wrong after time has gone into it. The owner of the higher-level question is often a person, not the model — ask them.

Where it comes from. Pólya’s How to Solve It (1945) recommends the move when a problem resists: drop part of the condition, generalize, solve an auxiliary problem. Gause and Weinberg’s Are Your Lights On? (1982) is built on the observation that a problem is a difference between things as desired and things as perceived, and that people solve the problem they were handed rather than the one they have.

An engineering example. Three rounds of asking a model to tune executor memory for a Spark job that keeps failing: each answer is reasonable, each changes a number, the job still dies on one task. One level up the question is “why does one task need so much memory?”, and the answer is a skewed key concentrating a quarter of the rows in one partition (92, 93) — a fix in the data, which no memory setting reaches.

Limits. “Two or three requests” is a rule of thumb, not a measured threshold. Going up has no natural stop — every problem is part of a strategy — so climb one level, restate, and act there. And some loops are legitimately iterative: polishing wording or tuning a parameter on measured results is convergence, not stagnation; the signal is understanding that stopped moving, not the number of turns.

Last technical review: 2026-08-04 · Reference · 112 principles