Briefing the machine is part of the craft
Good AI output starts with clear context, constraints, examples, and a human willing to edit the result.
The quickest way to get mediocre AI output is to treat the prompt like a vending machine.
Put in a vague request. Receive a shiny packet of average. Feel briefly impressed. Spend the rest of the afternoon cleaning up the mess.
Good AI work starts earlier than the prompt. It starts with the same thing good human work needs: context.
The brief is doing more work now
A strong brief has always mattered. AI makes that more obvious because the model will happily fill gaps you forgot to close.
If you do not describe the audience, it will choose one. If you do not name the tone, it will borrow the internet’s default voice, which sounds like a startup founder trapped inside a help article. If you do not explain the constraints, it will optimise for a world your project does not live in.
That is not the model being malicious. It is the brief being leaky.
The fix is not a longer prompt for the sake of it. The fix is a better brief: what this is for, who it serves, what already exists, what must not change, what good looks like, and how the result will be judged.
Constraints beat adjectives
“Make it premium” is not a useful instruction. “Use the existing type scale, keep the tone plain, avoid generic agency language, and do not add another call to action” is better.
Adjectives invite interpretation. Constraints shape behaviour.
For design work, that might mean naming the brand system, the spacing rules, the kind of motion that belongs, and the kind that does not. For code, it might mean naming the files, test commands, architectural patterns, and forbidden shortcuts. For writing, it might mean supplying the messy source material and saying which claims are known, which need checking, and which should be removed.
The model needs rails. So does the human reviewing it.
Examples carry taste
One of the most useful things you can give an AI tool is not a clever instruction. It is a good example.
Examples carry rhythm, judgement, pacing, and the small choices that are hard to specify directly. They show what “plain” means in this project. They show how much technical detail is appropriate. They show whether a paragraph should end with a tidy summary or a small turn of the knife.
This is why existing work matters. A model pointed at a real codebase, a real content archive, or a real brand guide has a better chance of staying inside the world of the project.
It can still drift. But at least it has something better than the average internet to drift from.
Ask for evidence, not confidence
AI tools are very good at sounding certain. That is not the same as being right.
I prefer prompts that ask for evidence. Name the files you inspected. List the commands you ran. Separate facts from assumptions. Flag claims that need checking. Explain why you chose this approach. Show the trade-off.
This slows the exchange down in the right way. It turns the output from a polished answer into something reviewable.
For production work, reviewable beats impressive.
Editing is not failure
There is a strange disappointment people sometimes feel when AI output needs editing, as if the tool has failed by not producing the final answer in one pass.
That is a bad standard. Most useful work needs editing. Human work definitely does. The question is not whether the first output is perfect. The question is whether it gets you to a better second pass faster.
A good AI workflow leaves the human with sharper material, not less responsibility.
Briefing the machine is part of the craft because the brief carries the judgement that the machine does not have. It names the context, narrows the field, protects the constraints, and gives the edit something to push against.
The prompt is not where craft disappears. It is one of the places craft shows up.
What a good brief saves is the non-AI version of this same lesson.