Seven habits for getting good work out of AI that do not expire with the next model, built on one idea: a model fills every gap you leave with the average, so leave fewer gaps.
- Be specific. State the goal, the audience, the length, what must not happen and what finished looks like, and add the details only you have.
- Check what matters. Ask for the reasoning, and verify names, dates, figures and citations, because confident text tells you nothing about accuracy.
- Edit the draft, not the prompt. When the output is close, say what is wrong with it instead of starting over.
- Keep the decisions and the private data. Hand over the labour but not the judgement, and do not paste in anything you would not want stored.
- The tool has no stake in being right. You carry the consequence, so you are the one who has to check.
A model fills every gap you leave with the average of everything it has seen. That single fact explains almost all bad output. Generic in, generic out — not because the tool is weak, but because you asked it to guess and it guessed the most common answer. Everything below is a way of leaving fewer gaps.
Seven habits
Most weak requests describe a topic and hope. Strong ones state the goal, the audience, the length, the thing that must not happen, and what "finished" looks like.
Real examples, actual numbers, your genuine position, the constraint that makes your situation unlike everyone else's. This is the whole difference between output that sounds like the internet and output that sounds like you.
When it writes for you (an email, a post, a plan in your voice): give it your real position. Without it, you get the average.
When it judges for you (is this idea good, is this draft right): keep your position back until it has answered, because models tend to agree with the person asking. Same advice, two jobs; the other half is on how AI influences you.Reasoning, September 2026 — reconciles this page with its companion page; the evidence that models lean toward the asker’s view is cited on how AI influences you.
Asking it to show how it got there gives you something you can inspect. An answer you cannot check is a claim you have to take on faith. Do not expect the request to make a reasoning model smarter, though: OpenAI says telling those models to think step by step can even hurt.OpenAI, Reasoning best practices, as quoted on the LLM cheat sheet, read at source 16 Sep 2026: “may not enhance performance (and can sometimes hinder it)”. Corrected 22 Sep 2026: this habit said asking for the steps “usually improves the answer”.
These systems produce fluent, confident text whether or not the claim underneath is true — and the confidence carries no information about accuracy. Names, dates, figures, citations, legal and medical specifics: check them, every time.
When something is mostly right, many people rewrite the prompt and start over. Faster: say what is wrong with what you got. "Second paragraph is too soft, cut the last sentence, keep the opening."
Before you start, know which part of the work is the part you are for. The judgement, the position, the taste, the responsibility for being right. Hand over the labour, keep the decisions — otherwise you become an editor of things you do not understand.
Other people's private information, credentials, anything under confidentiality, anything you would not want retained. Assume input may be stored or reviewed unless you have a written guarantee otherwise, and check what your specific tool and plan actually promise.
Before you send a request, fill the gaps yourself — the goal, the constraint, the specifics only you know — and check any consequential claim in what comes back.
Three ideas worth more than any template
Context beats phrasing. The industry called this "prompt engineering" and the name did damage — it implied secret words. The real skill is supplying the specifics only you have. Context is the craft.
Cheap output raises the value of judgement. When producing a draft costs nothing, the scarce thing becomes knowing which draft is good, and why. That is not a consolation — it is where the work moved. See the silver lining.
The tool has no stake in being right. It will produce a confident answer to a question it cannot answer, because that is what it was trained to do. You are the one who carries the consequence, so you are the one who has to care.
Spend your effort on context and judgement rather than clever wording, and treat a confident answer as unchecked until you have checked it.
WHY THIS PAGE IS SHORT
This site used to carry a separate section of prompting technique. Much of that kind of advice was true in 2024 and stale by 2026 — tied to quirks of models that no longer exist. Rather than maintain a collection of expired advice, it is condensed to the part that has not changed: supply context, keep judgement, verify claims. If a technique cannot survive a model release, it was never a skill.
Using AI is not cheating and it is not shameful — this site's own use of it is disclosed on its About page. What matters is disclosure and responsibility: say when a machine did the work, and stay accountable for whether it is true. The people who get hurt in this era are not the ones who used the tools; they are the ones who trusted output nobody checked.