Most disappointing AI output isn't the model's fault. It's the ask.
A vague question gets a vague answer, because a vague question genuinely is ambiguous. The model picks one reading out of many, and it's rarely the one you meant. Fixing that is the skill that pays back fastest, and it takes about twenty minutes to learn.
The core idea
A model has no idea who you are, what you're working on, or what a good answer would look like to you. It only has what's in front of it.
So the job isn't "phrase the question cleverly." It's close the gap between what you know and what the model can see.
Everything below is a specific way of doing that.
1. Say who it's for
This changes the answer more than any other single word you can add.
Explain vector databases.
versus
Explain vector databases to a product manager who needs to decide whether we need one. No code.
The first gets you an encyclopedia entry. The second gets you a decision aid. Same topic, completely different useful output.
2. Ask for the shape you want
If you're going to reformat the answer by hand afterwards, ask for that format up front instead.
- "As a table with columns X, Y and Z"
- "As five bullet points, one sentence each"
- "As a short email I can send to my team"
3. Give it the actual material
This is the single biggest lever you have: stop asking the model to guess at context it could just be given.
Instead of "write a follow-up email to a client who's gone quiet", paste the last email in the thread. Instead of "how do I improve this function", paste the function.
There is a limit, and Context Is Everything covers where more context starts making things worse. But almost everyone errs far on the side of too little.
4. Show an example when the format is unusual
For anything with a shape the model can't infer, one example beats three paragraphs of description:
Rewrite each line in this style:
Input: "The meeting is canceled."
Output: "Heads up — we're standing down the meeting."
Input: "The deadline moved to Friday."
Output:
This is sometimes called few-shot prompting. It's just showing your work.
5. Ask it to check itself, specifically
"Are you sure?" is close to useless; the model will either cave or dig in depending on how you asked, not on the claim.
Naming the failure mode at least gives you a list to check yourself:
Which claims above would you not be able to support with a source? List them.
When a prompt isn't working
Before rewriting it from scratch, work through this in order:
- Is the task actually clear to a human? Read your prompt as if someone else sent it to you. Could you do the task?
- Is the necessary information present? Nine times out of ten, no.
- Is it more than one task? Split it. Models handle one job at a time far better than four.
- Have you asked for a format? If not, you're inviting a wall of prose.
- Is it a task this tool is bad at? Arithmetic, current events, and anything needing a source are common dead ends. No prompt fixes those.
The short version
Say who it's for. Say what shape you want. Paste the actual material. Show an example if the format is odd. Then check the parts that matter.
That's most of it. The rest is practice.
This is the written version of the lesson The 5 Rules of a Good Prompt, which hasn't been recorded yet.