Prompt design is interface design
A prompt is a brief, not a command. What we can borrow from decades of UX when we talk to intelligent tools.
A prompt is closer to a brief than a command. You are not telling the system exactly what to do, step by step; you are describing the boundaries of a space and asking it to find something good inside them. Once you see it that way, prompting stops looking like programming and starts looking like interface design, a discipline with decades of hard-won lessons to borrow from.
Constraints describe the good answer
The most useful prompts spend their words on constraints, not commands. Who is this for. What is it part of. What does "good" look like, and what should be avoided. A vague request is an open door; the model will walk through it in whatever direction is statistically easiest, which is rarely the direction you wanted. A well-bounded request gives it room to be genuinely useful inside lines you chose deliberately.
There is a quiet paradox here that anyone who has designed a form or a product already knows: more constraint often produces more creativity, because the effort goes into the substance instead of the scope.
Context before instruction
The instinct is to lead with the task. The better habit is to lead with the context, then give the task. Tell the model who the reader is, what came before, what the piece sits inside, what success would feel like. One concrete example will do more than three sentences of adjectives, the same way a single mockup settles an argument that words keep circling.
Match the tool to the task.
Do not reach for your most powerful, most expensive model for everything. Match capability to what the task actually needs, reuse the stable part of a prompt, clear the context when you start something unrelated, and optimize for cost per successful outcome, not price per token.
A word on speed
If you work with these tools daily, three numbers are worth knowing, because different tasks care about different ones: time to first token (how responsive it feels), output tokens per second (how fast it sustains), and time to last token (the total wait). A chat assistant lives or dies on the first; a long generation on the last. Knowing which one your task depends on saves both money and patience.
Keep the judgment in the loop
The model proposes; you decide what ships. The most reliable habit I have is to make the answer wrong on purpose before trusting it: ask the model to argue against itself, to surface its assumptions, to find the weakest claim in its own reply. Most of the time it obliges, and the second pass is where the real thinking happens. Critical thinking is precisely the part that does not scale, which is exactly why it is worth protecting.
Good prompting, in the end, is not a bag of tricks. It is the same craft as any good interface: reduce ambiguity, respect the person's actual intent, make the easy things easy, and leave the consequential decisions to a human who is paying attention.
The AI answered you. Now make it wrong, before you make it final.