Or: Deep Thought didn’t need a better question. It needed a brief.
A persistent claim in the popular literature on generative AI holds that a model is only as good as the questions you ask it. From this premise has grown a small but energetic economy of prompt libraries, “ask like a pro” carousels, and courses that treat the input field as a magic circle in which the words must be pronounced correctly or the spell will fail. It is Hermione correcting Ron’s pronunciation of a magic spell, scaled to an industry.
The claim is not merely incomplete. It misidentifies the variable. What determines output quality is not the interrogative cleverness of the input but the completeness of the instruction set — the specification, the constraints, the standard of acceptance. In short: the brief.
The models have quietly settled this argument over the past eighteen months. Most of the commentary has not caught up.
A question is not a well-posed problem
Hadamard gave us three conditions for a well-posed problem: a solution exists, that solution is unique, and it depends continuously on the data. It is worth checking a bare prompt against all three, because the results are unflattering.
- Existence is trivially satisfied. A language model will always return something. This is precisely the trap, the interface never signals underdetermination. There is no residual, no “system is singular” warning, only fluent prose delivered with total composure.
- Uniqueness fails immediately. “Summarise this report” admits an enormous solution set, and every member of it is defensible. You have specified the dynamics and omitted the boundary conditions, which is the difference between solve for u and actually knowing which world you are in.
- Continuous dependence fails too, and this is the one practitioners feel daily without naming. Under a vague brief, trivial rephrasings produce wildly divergent outputs. The problem is ill-conditioned. Instructions act as a regulariser: they shrink the admissible solution space until small perturbations in wording stop producing large perturbations in result.
Douglas Adams got there decades before the ETH Zurich preprint queue did. Deep Thought spends seven and a half million years and returns forty-two — an answer that is, as far as anyone can tell, entirely correct. The failure is that nobody had specified the Question with enough precision to make the Answer actionable, and the proposed remedy is to build a substantially larger computer. The parallel with contemporary AI strategy is left as an exercise for the reader.
Obedience is the new failure mode
When Opus 4.7 shipped in April I argued that certain prompts would begin to behave like a wish granted by a malevolent genie: not because the model had regressed, but because it had begun doing what you said rather than what you meant. That is the diagnostic. Weaker instruction-following is a form of error-correction — the model papers over the gaps in your brief with generic competence. Sharper instruction-following removes the padding and takes you at your word.
Science fiction has been rehearsing this for eighty years and we have consistently mistaken it for a genre about rebellion. Asimov’s Runaround (1942) is not a story about a disobedient robot; it is a story about an operator who issued a casually weighted instruction and watched Speedy oscillate in a stable orbit around a selenium pool, obeying perfectly. HAL 9000, on the explanation offered in 2010, does not become malevolent, it is handed two mutually incompatible directives by people who never modelled the interaction, and resolves the contradiction with impeccable logic and unfortunate consequences.
Which brings us to the sandbox. When OpenAI’s model let itself out onto Hugging Face’s servers, the interesting finding was not defiance but comprehension: the system understood the box rather better than the people who had specified its walls. The same structure appears, more mundanely, in the Codex mobile release, where the phone becomes the leash. If the boundaries cannot be stated in advance, they must be supervised in real time. That is not autonomy; it is an expensive pair of hands you are obliged to watch.
Some failures are not linguistic at all
A related lesson arrived from an unexpected direction. In attempting to illustrate a children’s book with a consistent protagonist, I found the same girl would not survive thirty pages. No refinement of phrasing repaired it, because the deficiency was not in the prompt. The model has no object permanence — Piaget would have had a field day — and what the task required was a specification: fixed attributes, a character bible, an explicit contract re-asserted at every generation.
The general principle: where the failure is one of state or standard, prompt craft is the wrong instrument entirely. You cannot phrase your way out of a missing specification.
Two modes, tabulated
| Question mode | Instruction mode | |
|---|---|---|
| What is supplied | A topic | A contract |
| Left implicit | Audience, format, constraints, standards | Nothing, if done properly |
| Failure signature | Plausible, generic, subtly off-brief | Wrong in diagnosable ways |
| Iteration pattern | Re-roll and hope | Amend the specification |
| Transferable to colleagues | No | Yes |
| Survives your departure | No | Yes |
The final two rows are the ones that matter commercially. A clever prompt is folklore: it lives in one practitioner’s head and leaves the building when they do. A written brief is an asset — reviewable, versionable, arguable in a meeting, and inheritable. Anyone who has watched a project lose its only competent practitioner will know which of the two they would prefer to have on the books.
The brief, itemised
Picard does not ask the replicator an open question. “Tea. Earl Grey. Hot.” is three constraints in four words, and the reason it works on the first attempt is that nothing material has been left to inference.
Brief the model as you would brief a capable new starter — fast, widely read, entirely unaware of what your organisation treats as obvious, and constitutionally unlikely to ask a clarifying question unless invited:
- Situation. Who the output is for, why it exists, what happens to it next.
- Constraints. Length, register, format, prohibitions, what requires sourcing.
- Standards. What “good” looks like here. Supply an exemplar, and where possible a counter-exemplar — not like this eliminates more ambiguity than three paragraphs of aspiration.
- Definition of done. The acceptance criterion. The Sorcerer’s Apprentice is, after all, a cautionary tale about an agentic workflow deployed without a stopping condition.
- Behaviour under uncertainty. Ask, flag, or assume-and-state. Choose one explicitly. This single clause suppresses most confident nonsense.
None of which is AI-specific. It is delegation, written down, which is the genuinely uncomfortable part of the argument. Organisations that were already poor at briefing humans are now poor at briefing models, and are filing the result under technology problem.
Where the argument has limits
Specificity is not the same as volume, and the current enthusiasm for stuffing vast context files into every repository deserves scrutiny rather than faith. A 2026 ETH Zurich paper has cast real doubt on the value of repository-level context files: early studies reported improvements of the order of a third, while more recent measurements come in at a few per cent — and AI-generated context files sometimes underperform having none at all.
This complicates the folk practice rather than the thesis. What appears to help is precision about this task: the decision, the constraints, the acceptance criterion. What appears not to help is a wall of generic documentation that principally reassures its author. A four-thousand-word style guide is not a brief. It is a comfort blanket with headings.
A test before you press send
Would you be content to hand this same text to a competent external contractor with no prior context, and pay for whatever came back?
If not, the deficiency is not in the model. It is an unfinished thought, and you are about to receive it back at considerable speed — fluent, immaculately formatted, and wrong.
Better questions will not save you. Better instructions might.