PTN—10
Prompt Refinement
The system improves the request itself — clarifying, expanding, or rewriting — before spending effort on the answer.
Most people are bad at prompting and should never have to get good at it. Prompt refinement moves that expertise into the system: the AI asks a clarifying question when the request is ambiguous, expands a terse instruction into a fuller brief, or rewrites the prompt into what it will actually execute. Ten seconds spent on the request beats ten minutes spent on the wrong answer.
The pattern scales with stakes. For a cheap chat reply, guessing and letting the person correct is fine — clarifying questions on trivial requests are pure friction. For a twenty-minute research run or an image batch that costs real money, refining first is the only respectful design, which is why ChatGPT’s deep research asks scoping questions before it starts and image tools rewrite thin prompts into detailed ones.
Refinement must stay legible and consentable. When the system rewrites a prompt, show the rewrite and make it editable — a silent expansion that adds details the person never asked for is the system putting words in their mouth. The best implementations treat the refined prompt as a shared artifact: the person sees exactly what will run, adjusts it, and learns by example what a good request looks like.
The pattern emerged from an uncomfortable measurement: the gap between what models can do and what typical prompts ask of them, which early teams tried to close with prompt-engineering guides nobody read. Moving the expertise into the system was the correction. The tradeoff never disappears, though — every clarifying question spends the person’s momentum to save the machine’s effort, and the exchange rate flips with stakes. Maturity looks like ChatGPT’s deep research on one end and instant chat on the other: the same product interrogating a twenty-minute brief and answering a one-liner without a single question, because it has learned which is which.
When to use
- The downstream operation is expensive — long-running, costly, or irreversible — and a wrong guess wastes real effort.
- Requests arrive ambiguous in ways that materially change the output.
- There is a large gap between how novices phrase requests and what the model performs best on.
- You want people to improve at expressing intent without ever reading a prompting guide.
Design considerations
- 01Gate clarifying questions by stakes — ask before a research run, never before a one-line answer.
- 02Ask one round of questions at most, and make each answerable in a tap or a short phrase.
- 03Show any rewritten or expanded prompt before or alongside execution, and make it editable.
- 04Offer refinement as an affordance — an “improve this prompt” action — not only as an interruption.
- 05Carry the person’s original wording forward; the refined version augments intent, never replaces it.
- 06Let people skip refinement and run the raw request when they know exactly what they want.
Pitfalls
- ✕The interrogation: a wall of clarifying questions before trivial requests, adding a form to what should be a sentence.
- ✕Silent rewriting — the system expands the prompt behind the scenes and generates something the person never asked for, with no way to see why.
- ✕Refinement that flattens intent: the rewrite normalizes a deliberately unusual request into the generic version of itself.
- ✕Asking questions the answer would not change — clarification as ritual rather than as information the system will actually use.
In the wild
- ChatGPT
- Deep research asks scoping questions — angle, depth, constraints — before committing minutes of work, refining the brief while the refinement is still cheap.
- Midjourney
- Sparse prompts are expanded into richer descriptions before generation, trading literal fidelity for output quality on thin input.
- Claude
- The Claude Code planning flow restates an ambiguous task as an explicit plan for approval — refinement of intent before execution against it.
- Perplexity
- Related and follow-up question suggestions after each answer teach iterative narrowing — refinement distributed across the conversation.