How to write a prompt that works the first time

"Clean up this customer list."

That's a perfectly normal thing to type into an AI, and it's a trap. The AI will do what you asked. It'll remove duplicates, fix the capital letters, and quite possibly delete every customer who hasn't ordered in a while, because to a machine with no context an inactive customer looks a lot like clutter. Maybe those were people you wanted to keep. You just didn't want them in the next mailing.

Nothing about that answer would be wrong. It's a good answer to a bad question. (I borrowed this example from a prompting video, because it's the cleanest one I've seen.) Most of the "the AI is dumb" moments I've had came down to the same gap: I told it what, and I kept the why in my head.

Say what, then say why

Here's the shape I use now for almost everything. It has three parts and the middle one does most of the work.

What: <the task, one short line>
Why: <the reason, plus the real limits: what good looks like, what to leave out, the deadline>
Pause: <when to stop and ask me>

That customer list, rewritten:

Prep my customer list for Monday's move to the new billing system.

Why: only customers active in the last 90 days are moving over. Dedupe by
email. Anyone inactive gets archived, not deleted. I may want them later.

Pause and ask me before deleting anything, or if a row looks wrong and
you're not sure what to do with it.

The second version isn't much longer. It's just honest about what's going on. Once the AI knows the list is for a billing move, it can make the dozen small calls I didn't think to write down (keep the phone numbers, don't touch the notes column, flag that one customer with two emails). It makes them the way I would have, because now it knows what I'm trying to do.

The framing comes from that same video, which walked through Anthropic's own prompting advice. It stuck with me enough that I set my tools up to rebuild every request I type into this shape before anything runs. I don't trust myself to remember the why when I'm in a hurry, so the setup remembers for me. The newer models are good at using intent. Give them the reason and they connect it to the right details on their own.

Keep it shorter than you think

This one surprised me. Once you learn prompts matter, the instinct is to write longer ones. Rules, roles, formatting demands, "you are a world-class expert in..." the works.

The advice from that video was the opposite, and it matches what I've seen since: over-specifying makes the answers worse. Not a lot worse, but stiffer. A prompt stacked with fifteen rules makes the model spend its effort obeying the rules instead of thinking about the problem. It's a bit like handing a good barista a laminated card that covers the temperature, the grind and the exact pour angle. They'll follow it. You'll get a fine coffee. You won't get their best one.

So now I write the what, the why, and any limit that truly matters, and then I stop. If the answer misses, I add one line and try again. I don't front-load ten.

The pause rule is the part people skip

The assistants we have now will keep going. Ask for help with an email and some of them will happily draft the email, a follow-up, and a reply to a reply you never got. That's fine for an email. It's not fine when the task touches something you can't undo.

A pause rule is one sentence: when should it stop and check with you? Mine is almost always some version of this:

Stop and ask me only if the work truly needs my input: something that
can't be undone, a real change in scope, or something you're unsure about.
Otherwise keep going.

The "otherwise keep going" half matters as much as the first half. Without it, some tools ask permission for every tiny step, which is its own kind of useless.

Three quick before/afters

A birthday message for a coworker:

Before: write a birthday message for my coworker

After: Write a short birthday note for Dana, a coworker I like but don't
know well. Why: it's going in a group card, so it should be warm but not
personal. Two sentences, no jokes about age.

A cover letter:

Before: write a cover letter for this job [paste]

After: Write a cover letter for the job below. Why: I'm switching from
retail management into operations, so the letter has to connect running a
store to running processes. Use the two examples in my notes. Under 250
words. Ask me if you need a detail you can't find.

A recipe:

Before: give me a dinner recipe

After: Give me one weeknight dinner for four. Why: two of us don't eat
pork, one kid hates anything spicy, and I have about 30 minutes after work.
Use what's in my list below first.

Notice none of the "after" versions are clever. They just stop making the AI guess.

When it still misses

It will, sometimes. When that happens I don't start over. I reply in the same chat with the one thing that was off ("shorter," "less formal," "you deleted the inactive people, put them back and archive them"). Starting fresh throws away everything it just learned about what you want.

And if you'd rather not write prompts from scratch at all, that's fine too. The free prompts on NeLocker already have the what and the why built in. Open one, fill in your details, and you're most of the way there. If you want the longer, step-by-step version of this, the beginner's guide to writing prompts covers it, and the Claude guide has tips for that model in particular.

Quick answers

What is the most important part of an AI prompt? The reason behind the request. When you tell the AI why you want something, it can make the dozen small choices you didn't spell out, and it makes them the way you would.

How long should a prompt be? Usually a few sentences. Long, rule-heavy prompts tend to box the model in and get stiffer answers. Say what you want, why, and any hard limits, then stop.

What is a pause rule in a prompt? A line that tells the AI when to stop and check with you, for example before anything that can't be undone or when it's unsure. Without one, a capable assistant decides for itself how far to go.