How to write a prompt the AI actually understands

You type something into Claude or ChatGPT. The answer comes back close, but not quite right. You rewrite the same request three times, get three slightly different wrong answers, and start wondering if the tool is broken.

It isn't. The model did exactly what it was told. The instruction just wasn't clear enough to be read only one way.

The instruction comes first

Put what you want at the top of the message, then the material it should act on, separated clearly, a line of quotes works fine. A model reading a wall of context followed by a buried instruction has to guess which part matters. Lead with the ask.

Replace vague words with numbers

"Keep it short" means something different every time you type it. "3 sentences" or "under 200 words" means the same thing every time. Any word that could mean five different lengths, tones, or formats should get replaced with something you could actually measure.

Say what to do, not what to skip

"Don't ask for personal details" leaves everything else open, including asking in a different way. "Refer them to the help article instead" gives the model somewhere to go. A list of don'ts is a fence with no gate. One instruction pointing forward works better than five pointing away.

Show it, don't explain it

If you want a specific format, a table, a list, a particular tone, one example of that shape does more work than a paragraph describing it. Models are pattern matchers first. Showing beats telling almost every time.

Start simple

Write the plainest version of the request. See what comes back. Add detail only where it actually went wrong. A prompt built to cover every edge case before you've even tried it once usually takes longer to write and works no better than one tightened after a single round of feedback.

If you're using Claude specifically

Tell it why you're asking, not just what. "I'm writing this for X, they need Y, with that in mind: [the request]." Claude connects the dots itself once it understands the reason behind the ask. It's a small addition and it changes the output more than most people expect.

One more: say whether you want it to think or act. Claude can be capable enough to start fixing something you were only describing out loud. If you just want an assessment, say so. If you want it to act, say that instead.

It's the same skill as writing a clear brief for a client, or a spec for a developer: removing every word that could be read two ways. You've done this before. You just haven't had to do it for a machine reading this literally.

Copy this into your AI's custom instructions

Most AI tools, ChatGPT, Claude, Gemini, let you save a standing instruction that applies to every chat. Paste this in once and it will apply the rules above automatically, without you having to remember them each time.

Before responding to any request I give you, check it against these points. If something's unclear, ask me one question or suggest a tightened version before you proceed:

1. Is the scope specific? If I used a vague word like "short," "professional," or "a few," ask for something measurable instead, a word count, a sentence count, a format.
2. Is the instruction positive? If I told you what not to do, check whether I also said what to do instead. If not, ask.
3. Do I want a particular format? If so, ask me to show an example rather than relying on a description.
4. Am I asking you to think or to act? If it's not clear whether I want analysis or action, ask before doing either.
5. Is there a reason behind the request? If I haven't said who or what it's for, feel free to ask, it changes how you should answer.

Once the request is clear, lead your answer with the outcome first, then the supporting detail.