The five rules are about writing one good prompt. This is a level up: four shapes you can reach for, so you are picking a known pattern rather than starting from a blank box every time.
There are only four worth memorizing. One of them has quietly stopped being necessary, and knowing that saves you more time than the other three combined.
1. Role
Open by telling the model what position it is answering from.
You are reviewing this contract as the person who will have to
deliver the work, not as the lawyer who wrote it.
[contract]
What is going to be a problem in month three?
What it actually does. It is not method acting and the model does not "become" anything. It narrows which patterns are most likely to continue the text. "As a lawyer" and "as the person delivering the work" pull toward different vocabulary, different concerns, different things flagged as risky.
Where it is overrated. "You are a world-class expert" does close to nothing. There is no reservoir of superior answers that flattery unlocks. The useful part of a role is never the seniority, it is the vantage point: whose interests, which constraints, what they would notice. Swap "you are an expert marketer" for "you are the person who has to answer the support tickets this campaign generates" and the answer changes completely.
2. Examples
Show one or two instances of what you want instead of describing it.
Rewrite these in the same style:
Input: "The Q3 numbers came in below projection."
Output: "Q3 missed. Here is what we are changing."
Input: "We are experiencing higher than normal support volume."
Output:
When examples beat instructions. Any time the thing you want is easier to recognize than to define. Tone, formatting, level of detail, house style. Three paragraphs describing your voice will lose to two examples of it every time.
When they do not. Examples anchor hard, and that cuts both ways. If your two examples happen to share something you did not intend, both short, both starting with a verb, both about bad news, the model will treat that as the rule. If you need range, show range. And for a task with one correct answer rather than a style, examples add tokens without adding much.
3. Reasoning, which you probably no longer need to ask for
For years the advice was to append "think step by step". It worked, and the mechanism was real: reasoning tokens are more text in the window, and every subsequent token conditions on them, so the model gets to do work before committing to an answer.
That advice is now mostly obsolete, for a specific reason. Current frontier models do this internally by default. Asking for it explicitly is largely redundant, and on some tasks it produces a performance of reasoning bolted on top of reasoning that already happened.
What replaced it is a dial rather than a phrase. The practical question is no longer whether the model reasons, it is how much you want it to spend on this particular question, and most tools now expose that directly as a thinking or effort setting.
Two places the old technique still earns its place:
- Smaller and cheaper models. Few-shot reasoning examples still measurably help models below the frontier tier. If you are using a fast, cheap model for volume work, keep it.
- When you want the reasoning as output. "Show your working" is different from "think step by step". You are asking for the trace because you intend to read it and check the logic, which is a legitimate and different request.
This is the pattern to know about mainly so you stop spending prompt space on it out of habit.
4. Critique
Ask for the work, then ask the model to attack it.
Here is a draft announcement. Before you improve it, list the three
things a skeptical reader would push back on hardest, with the exact
sentence that would trigger each one.
Why this one is the highest leverage of the four. Everything a model generates comes out at the same confident register, because fluency and accuracy come from the same process. Critique is the only pattern here that puts something in front of that.
How to do it so it is not theater. A bare "is this good?" gets you agreement, because agreement is the most probable continuation of a question shaped like that. Three things make the difference:
- Name the critic. "A skeptical reader", "the person who has to maintain this", "someone who disagrees with the premise". A vantage point again.
- Demand a number. "The three weakest claims" forces a ranking. "Any problems?" invites "looks good".
- Make it point at specifics. Ask which exact sentence, which exact line. Vague criticism is as cheap to generate as vague praise.
Best used as a separate turn, or a separate conversation entirely. A model asked to produce and critique in one breath tends to produce something it can then congratulate.
Combining them
They stack, and the useful stack is usually role plus examples for the first draft, then critique in a second message. That is two patterns doing the work and one doing the checking, which is roughly the right ratio.
What does not stack well is four of them at once. A prompt carrying a role, six examples, a reasoning instruction and a self-critique clause is mostly scaffolding, and the ask gets lost in it.
The short version
Role works when it supplies a vantage point, not a job title. Examples win whenever the target is easier to recognize than to describe, and they anchor harder than you expect. Reasoning is now built in, so stop asking for it and start deciding how much to spend. Critique is the one that catches errors, and only if you name a critic, demand a count, and ask for specific lines.
Four shapes. Almost everything else is a variation.