How to improve an AI prompt
A weak prompt and a strong prompt are usually only two sentences apart. Here are the five edits that close the gap — each one shown as a rewrite you can drag through.
Nearly every disappointing AI answer traces back to the prompt, not the model. And it is almost never a matter of finding magic words — it is a matter of supplying information the model never had. Below are the five edits that account for most of the difference, each shown as a before and after you can drag through. Grab the divider and pull.
1. Give the model a role
Without a role, a model answers as a generic assistant: correct, cautious, and bland. A role does not make it smarter — it narrows the range of answers it considers plausible, which is what you actually want.
Be specific about the seniority and the context, not just the job title. "A lawyer" is weaker than "a contract lawyer reviewing a freelance agreement for a client who cannot afford a dispute."
With a role
You are a copywriter for a performance running brand. Your readers are marathon runners who already own three pairs of shoes and are sceptical of marketing claims. Write a product description for our new running shoe.
Without a role
Write a product description for our new running shoe.
2. Say what you want done, not what the topic is
A topic is not a task. "Something about our Q3 numbers" gives the model no verb to act on, so it guesses — usually a summary, because a summary is the safest guess.
Name the verb: analyse, compare, rank, rewrite, critique, extract. If you want more than one thing, number them, and the model will keep them apart.
A task
Compare our Q3 sales numbers against Q2 and do three things: (1) name the two largest changes, (2) suggest the most plausible cause for each, (3) flag any figure that looks like a data-entry error.
A topic
Our Q3 sales numbers.
3. Define the output format
This is the single most skipped edit, and the one that saves the most time — because the reformatting you would have done by hand is work the model can do for free.
Describe the shape you want: the sections, the length, the order, whether it is a table or prose. If you have to paste the answer somewhere specific, say where.
Format defined
Summarise this meeting transcript as: a one-sentence outcome, then a bullet list of decisions, then a table of action items with columns Owner | Task | Due date. Anything discussed but not decided goes under a final heading called "Open".
No format
Summarise this meeting transcript.
4. Add constraints — especially negative ones
Constraints are what stop a technically correct answer from being useless. Length, audience, tone, and reading level all belong here.
Negative constraints are the underrated half. Telling the model what not to do ("no analogies", "do not invent figures", "if a fact is not in the source, say so") removes the failure modes you would otherwise have to edit out by hand.
Constrained
Explain our pricing change to existing customers in under 120 words, at a reading level a busy non-technical buyer can skim. Do not apologise, do not use the word "unfortunately", and do not invent a reason we have not given you. If the reason is unclear from my notes, leave a bracketed [REASON] placeholder instead of guessing.
Unconstrained
Explain our pricing change to customers.
5. Turn the moving parts into placeholders
Once a prompt works, the parts you will change next time are almost always nouns: the recipient, the topic, the tone, the language. Everything around them is scaffolding you built once and should never type again.
Marking those nouns as placeholders is what turns a good one-off prompt into a template. It is also the cheapest edit on this list, because you are not writing anything new — you are just noticing which words move.
Reusable template
You are a support lead. Write a {{tone}} reply to {{recipient}} about {{issue}}, under {{word_limit}} words. Do not promise a date we have not confirmed.
One-off prompt
You are a support lead. Write a friendly reply to Mr Smith about his delayed invoice, under 100 words.
Putting the five together
Applied at once, the five edits turn a one-line request into a short brief — and a short brief is what the model needed from the start. Role, task, format, constraints, placeholders: if an answer disappoints you, one of those five is almost always missing.
The catch is that a good prompt takes a few minutes to write, which only pays off if you never write it twice. That is the whole reason TextDeck exists: you keep the brief, fill in the placeholders, and copy the result. It is free, it runs on macOS, iOS and Android, and the prompt quality score in the app checks these same five pillars while you type.