Once you have the basics of prompting down, the next step is not a longer list of tricks. It is a small set of reliable patterns, reusable moves that keep working on real tasks, under pressure, across model versions. These are the ones I actually reach for, stripped of the hype.
The role, the task, the format
The most durable structure is boring and it works everywhere: one line of who the model should be, a clear statement of the task, and a precise description of the output you want. "You are an editor. Tighten this paragraph to under 80 words without losing the main point. Return only the edited text." Role, task, format. It is not clever. It is reliable, which is better.
The one good example
When the output is hard to describe, show it instead. A single example of exactly what you want, its length, tone, and shape, steers the model harder than a paragraph of adjectives. This pattern survives model changes because it does not depend on any trick. It just gives the mimic something to mimic.
Describing the output is guessing. Showing one example is instructing. Instructing wins.
Think, then answer
For anything with reasoning, ask the model to work through it before committing. "Reason step by step, then give the final answer." On hard problems this genuinely improves accuracy, because it stops the model from blurting a confident first guess. On modern reasoning models it happens automatically, but on many tasks it is still worth requesting explicitly.
The critique-and-revise loop
One of the highest-value patterns: have the model produce an answer, then ask it to critique its own work against specific criteria, then revise. "Now list three weaknesses in that draft, then rewrite it fixing them." You get a second, better pass for the cost of one extra message, and it catches things a single shot misses.
The constraint-first prompt
State the hard limits up front, not as an afterthought. Length, what to include, what to avoid, the format. Models honor constraints they see early far more reliably than ones tacked on at the end. Put the guardrails before the request, not after.
The takeaway
Forget the magic phrases. These five patterns, role-task-format, one good example, think-then-answer, critique-and-revise, and constraint-first, are the reusable machinery of good prompting. They work across models, they survive updates, and they degrade gracefully. Learn them once and you stop chasing the trick of the week, because you are building briefs, not casting spells.