Prompting is a design skill

Every week someone sends me a screenshot of a "god-tier prompt" — six hundred words of role-play, threats about tipping, and instructions to think step by step. Then they ask why their output still looks generic. The prompt isn't the problem. The thinking underneath it is.

The words are the last five percent

When a client's AI output is bad, it's almost never because the phrasing was wrong. It's because nobody defined what "good" looks like before they started typing. That's not a language problem — it's a design problem, and designers have been solving it for decades under a different name: the brief.

A model can only aim at a target you've described. If your target is "write me a good caption," you've handed it a blank map and asked it to find treasure. You'll get the statistical average of every caption on the internet. Which is, by definition, generic.

What actually changes the output

Four things, roughly in order of impact:

Context. What the model can see. Your brand guide, past work, the actual audience, real examples of what you liked and what you didn't.
Constraints. What it can't do. Length, tone, forbidden words, format. Constraints are where taste lives.
Examples. Two or three samples of the output you want beats two paragraphs describing it. Show, don't tell — same rule as everywhere else.
Iteration. First output is a draft, not a delivery. The skill is knowing what's wrong with it and saying so precisely.

Notice that only the last one involves typing clever things. The other three are preparation — the same preparation you'd do before briefing a junior designer.

If you can't explain what you want to a competent human in one paragraph, you can't explain it to a model either. Fix the paragraph first.

A structure that holds up

I use roughly the same skeleton for anything non-trivial. Not because it's magic, but because it forces me to answer the questions I'd otherwise skip:

ROLE      — who is speaking, and to whom
CONTEXT   — the brand, the audience, the situation
TASK      — one clear job, not five
CONSTRAINTS — length, tone, format, what to avoid
EXAMPLES  — 2-3 samples of "right"
OUTPUT    — exactly what shape you want back

The value isn't the template. It's that filling it in honestly takes about four minutes, and those four minutes save the forty you'd spend regenerating.

Where people go wrong

The most common mistake I see is asking for too much at once. "Write the landing page" produces mush. "Write six headline options for a landing page selling X to Y, under nine words each, no exclamation marks" produces something you can work with. Narrow the task and the quality jumps immediately.

The second mistake is accepting the first output because it's fast. Speed is only valuable if the thing is good. A bad deliverable produced in ten seconds is still a bad deliverable.

Why this matters if you're hiring

People sometimes assume AI means the work is now trivial, so it should be nearly free. In practice it moved the effort, it didn't remove it. The typing got fast. The judgement — knowing what to ask for, spotting what's wrong, deciding what ships — is still the entire job, and it's still the part you're paying for.

That's roughly what I do all day: hold the standard while the machine does the volume.