PMaker home
Search
中EN
System
Color theme
空格的键盘
空格的键盘空格的键盘产品经理 · AI Builder
Four tricks, four mechanisms — understand the principle and you'll know when to use eachHigh-leverage tricksEach treats a different problemAsk firstWhen your requirements are fuzzy, have it ask 3 to 5 questions before startingAnchor to a referencePoint to a concrete reference instead of using adjectivesThink before answeringWrite out the reasoning steps explicitly for extra chances to correctRole framingShifts tone and wording, not capability or accuracyThese aren't magic spells — they work only when the scenario fits

You don't need many tricks — these four cover most scenarios. The point is knowing what each one treats.

A Few High-Leverage Tricks

Asking questions first, reference anchoring, think-before-answering — and what role prompts really do.

Online lists of "prompting tricks" run to dozens of items, most of them the same idea restated. Only four are worth remembering — and each has a specific scenario. They are not magic spells.

Symptoms you'll recognize:

  • You haven't fully thought through what you want, so the prompt is vague too.
  • You add "you're a senior expert," it sounds more professional, and accuracy doesn't move.
  • Complex tasks fail somewhere in the middle, then cascade to the end.

Ask first

The most counterintuitive and most useful trick. When you can't articulate what you want, don't force the writing — let it ask you.

The move is simple: add one line to the task — "Before you start, ask me the 3 to 5 most critical questions, and begin once I've answered."

Why it works: it converts "you must conjure up all the background from nothing" into "you only answer a few concrete questions." Much easier. And the questions it asks tend to be exactly the prompt parts you forgot.

There's a deeper bonus: your answers become part of the context. So this trick is really letting the model complete the prompt for you — freeze the whole exchange afterward and you have a reusable prompt.

Where it fits: fuzzy open-ended requests, first time doing a task, drafting a complex document. Where it doesn't: batch automation — there you want determinism, not a conversation.

Anchor to a reference

This extends the few-shot idea: when you can point at a concrete reference, don't reach for adjectives.

"Make it more concise" is weaker than "match the information density of this page." "Sound more professional" is weaker than "write in the voice of our existing help-center articles." "Make it more modern" is weaker than pasting a design.

The reason is the same one from earlier: adjectives are compressed information, everyone decompresses them differently, and the model averages the guesses. A concrete reference carries the full detail.

One catch: say which aspect to reference. If you paste an article and say "reference this," it won't know whether you mean structure, tone, length, or angle. Say "match its paragraph length and subheading style, but the content is entirely different" — and results get far more stable.

Think before answering

Add one line to the prompt — "list your analysis steps first, then give the conclusion." For multi-step tasks, accuracy usually rises noticeably.

The mechanism was covered in the reasoning-models article: the model's thinking is its output; writing more is computing more. When intermediate steps are written out explicitly, every later prediction sees the earlier reasoning — several extra chances to catch a mistake. The chain-of-thought paper verified exactly that: making the reasoning explicit reliably improves accuracy on hard tasks. chain-of-thought

Three cautions:

1. It costs something. Intermediate steps are output tokens too — slower and pricier. Skip it for simple tasks.

2. Don't add it if you're already on a reasoning model. It's doing this internally; the extra request can actually interfere.

3. Ask for process and conclusion separately. Reasoning in one block, conclusion in another, so your program can take only the conclusion and never show the thinking to users.

A variant worth knowing: have it self-check before delivering. "After your answer, review it against each constraint above and fix any mismatches." This shines when there are hard constraints — length, banned words, required elements.

The truth about role framing

"You are a senior lawyer with twenty years of experience" opens nearly every prompt template. It does something real, but its effect is badly overstated — worth getting straight.

What it actually does: shift the distribution of vocabulary and tone. Call it a lawyer and it reaches for legal phrasing, more careful sentence construction, standard disclaimers. That's a direct consequence of the narrowed probability distribution — the words near "lawyer" in the corpus genuinely differ from those near "elementary school teacher."

What it does not do: make the model know more law. The model's knowledge was fixed at training time. A role line adds no knowledge and raises no factual accuracy.

This distinction matters in products. Teams assume adding "you're a senior medical advisor" makes output more reliable, then relax their review — a dangerous misunderstanding. The tone gets more professional while the error rate stays the same, and because it sounds more authoritative, the errors become harder to spot.

So use it correctly: treat role framing as a tone and style switch, not a capability switch. For accuracy, provide material, give examples, and keep a human review at the end. Anthropic's guide positions it the same way — persona sets style and context; it doesn't fake capability. anthropic-prompt-eng

And rather than a vague "senior expert," write concrete behavioral constraints: "always cite which section of the material each claim comes from." "When unsure, say so explicitly; don't guess." These outperform any title.

References

  1. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
  2. Prompt engineering overview — Anthropic