Photo by Glenn Carstens-Peters on Unsplash
Prompt guides tend to sell frameworks with initials. They are not wrong, and they are not where the improvement comes from either. Most of the gap between a useless answer and a useful one is explained by three things, only one of which any acronym covers.
One: Supply the Material
The single largest lever, and the least discussed. These models are far better at transforming text you provide than at producing text from nothing.
“Write a project update” produces generic filler because you gave it nothing to work with. Paste your scrappy bullet points, last week’s update, the meeting notes, and the same request produces something specific, because the specifics came from you.
Two: Name the Reader
Audience changes vocabulary, length, structure and what gets assumed. It costs six words and it changes everything downstream.
| Vague | Specified |
|---|---|
| Explain our refund policy | Explain our refund policy to a frustrated customer whose claim we are declining |
| Summarise this report | Summarise this report for a director who will read three sentences before deciding |
| Make this clearer | Make this clear to someone who has never used the product |
Three: Say What Good Looks Like
Not just the task but the shape of an acceptable answer. Length, format, what to include, what to leave out. Models default to hedged, comprehensive and medium-length because that is the safest average, and the average is rarely what you wanted.
“Three options, one line each, no explanation” gets you three options with one line each. “Give me some options” gets you a lecture.
Photo by Emiliano Vittoriosi on Unsplash
If you could not answer the question yourself given the information in your prompt, neither can the model without inventing.
What Does Not Help As Much As Advertised
- Assigning an expert persona. “Act as a senior lawyer” mostly changes tone. It does not add knowledge and it can add unearned confidence, which is worse than a plain answer.
- Politeness. Please and thank you cost nothing and do nothing. Keep them if you like them.
- Elaborate templates. Long scaffolds copied from prompt libraries add tokens more often than they add quality. The three things above cover most of it.
- Asking it to be accurate. “Do not make anything up” does not prevent fabrication, because the model cannot tell when it is doing it.
The Follow-Up Is Where the Work Happens
Treat the first response as a draft you are editing, not a result you are grading. The conversation already holds the context, so corrections are cheap.
Two follow-ups worth having ready. “What would someone who disagreed say?” counteracts the default agreeableness, which is stronger than people expect: push back on a correct answer and you will often get a retraction and a worse one. And “which parts of that are you least sure about?” surfaces the weak joints faster than checking every claim.
Photo by Mike Tinnion on Unsplash
A citation is a claim the model made, not a receipt.
The Habit That Matters Most
Ask for sources you can open, then open them. Fabricated references are a documented failure of these systems, and the reflex to click is the difference between using one well and repeating its errors in public.
It is not a niche concern. The Tow Center at Columbia tested eight AI search tools across 1,600 queries and found source attribution wrong more than 60% of the time, with the best performer still around 37%. A citation is a claim the model made, not a receipt.
Putting It Together
A prompt that works is usually unglamorous: here is my material, here is who it is for, here is what a good answer looks like. Then one or two follow-ups to sharpen it. No acronym required, and it takes less typing than the frameworks do.
Where this lands in practice, including the settings worth checking first, is in How to Use ChatGPT.
FAQ: Frequently Asked Questions
What makes a good AI prompt?
Supplying your own material, naming the reader, and describing what an acceptable answer looks like. Those three account for most of the difference; the rest is refinement in follow-ups.
Do persona prompts work?
They adjust tone rather than capability. “Act as an expert” does not add knowledge, and it can produce more confident phrasing without more accuracy.
How long should a prompt be?
Long enough to include your material and your constraints, no longer. Added length only helps when it adds information the model did not otherwise have.
Can I stop it inventing facts?
Not by asking. Reduce exposure by supplying the source material yourself and by requiring links you can open for anything you intend to repeat.
Model behaviour changes between versions and the tendencies described here are general rather than fixed. The research cited tested tools available at publication.