The most expensive prompt in most companies is three words long: “make it better.” It is cheap in tokens and costly in people. Because it hands the model a judgment you have not made, you get the model’s generic idea of better, and then you spend several rounds steering it toward yours. Replace it with a short brief that says who the work is for, what is wrong with it now and what good looks like.
What “better” means to a model
Ask a model to improve a text with no target and it does what an eager new hire would do: a little of everything. Try it on any draft and you will usually get a longer text, a more formal register, a new subject line, an upbeat opening and a call to action at the end. Some of that may help. You cannot tell which change was the improvement, and neither can the model, because nobody said what was wrong.
“Make it more engaging,” “punch it up,” “make it sound professional” and “fix this” all fail the same way. Each names a direction and no destination.
Where the cost lands
Rewriting a 300-word email costs a few cents at most on any current model, so the tokens hardly matter. What costs money is the loop:
You: Make it better. Model: (rewrites everything, and it is longer) You: Shorter. And keep the line about the price change. Model: (shorter, price line back, a new opening) You: Less salesy. Our customers hate exclamation marks. Model: (calmer, but the deadline has disappeared) You: Put the March 1 deadline back in the first paragraph.
Four rounds, and every instruction in rounds two to four was something you knew before round one. That was the brief. It just arrived in installments.
A 300-word email is about 400 tokens. Monthly cost of 1,000 calls with about 400 tokens in and 400 out each. List prices checked October 20, 2026; your bill may differ.
The worse case is when nobody runs the loop. A draft that looks polished goes out, and whoever receives it has to work out what it means. Researchers at BetterUp Labs and the Stanford Social Media Lab call this workslop: AI-generated work that looks finished but does not move the task forward. In their survey of US desk workers, reported in Harvard Business Review in September 2025, about 40% said they had received some in the previous month, and they spent close to two hours dealing with each instance. The person who typed “make it better” saved a minute. The reader paid for it.
- of US desk workers surveyed had received workslop in the previous month
- About 40%
- spent dealing with each instance
- Close to 2 hours
Source: BetterUp Labs and Stanford Social Media Lab, reported in Harvard Business Review, September 2025.
If you manage people who use AI, here is a cheap habit: when someone sends you an AI-drafted document, ask to see the prompt. A one-line prompt is the earliest warning of workslop you will get.
Write the critique first
Before you ask for a rewrite, write one sentence about what is wrong. That sentence is most of the brief.
- Instead of “make it better”: “The second paragraph buries the price change. Customers on the old plan need their new price and the date in the first two lines.”
- Instead of “make it more engaging”: “Our readers are finance managers who skim. Cut the anecdote and lead with the number.”
- Instead of “make it sound professional”: “This goes to a hospital procurement team. Remove the jokes and the exclamation marks, and keep every figure.”
If you cannot write that sentence, the draft may be fine. Or you have not read it closely enough yet to hand the job to anyone, person or model.
What a good brief contains
A brief for a model holds the same things as a brief for a good freelancer. Our guide to the ten checks of a well-written prompt covers each one in depth. For a rewrite, they come in roughly this order:
- The task, with a precise verb: rewrite, shorten, restructure, translate, check.
- The reader and the goal. A renewal email that should get a reply is a different text from one that should prevent a support ticket.
- The material, fenced off from your instructions: the draft, plus any facts the new version must carry.
- What stays and what goes. The price, the date and the legal line stay. The anecdote and the second example can go.
- The shape: length, format, whether it needs a subject line.
- The limits: no exclamation marks, no claims we cannot back up, American English.
- What good looks like: two sentences from a past email you liked, or a plain description of the standard.
- What to do when unsure. “If a fact is missing, write [MISSING] instead of guessing” is the cheapest defense against a confident invented detail.
The tenth test, fit, is about laying the prompt out for the model that will read it. The Prompt Grader checks all ten as you type, for free, and with your own key it can rewrite the prompt for Claude, GPT or Gemini.
A brief, before and after
Here is the same job written both ways. The product, prices and dates are made up.
Make this email better. [pasted draft]
Rewrite the email below for customers on our
Starter plan. They need to understand that their
price rises from $12 to $15 a month on March 1,
and that switching to annual billing keeps $12.
We want fewer "why was I charged more" tickets.
Keep: the new price, the March 1 date, the annual
option and the link. Cut: the roadmap paragraph.
Format: a subject line, then at most 120 words.
Plain and friendly, no exclamation marks,
American English.
If a fact you need is missing, write [MISSING]
instead of guessing.
<draft>
{{draft}}
</draft>The brief is longer by one paragraph. It names the reader, the change they must understand, the outcome the company wants, what must survive the edit, the length and the tone. The model no longer has to guess any of those, so it stops guessing. The first draft is usually close, and the second round, if you need one, is about wording instead of direction.
It is also reusable. Next quarter, only the numbers change.
Reuse the briefs that work
When the same brief comes up every month (price change emails, release notes, job posts), turn it into a template. Keep the fixed parts and replace each detail that changes with a variable such as {{new_price}}. Our prompt templates are built that way, for jobs from support replies to system prompts, and each scores at least 8 of 10 on the free score.
Then measure the brief instead of admiring it. Run it on five real drafts in Prompt Tester with a check for each must-keep fact (the price, the date, the link), and you will know whether it works before your team relies on it.
- 1Write what is wrongOne sentence. It is most of the brief.
- 2Write the briefReader, goal, what stays, shape, limits, and what to do when unsure.
- 3Save repeats as templatesKeep the fixed parts; only the details change.
- 4Test it on five real draftsWith a check for each must-keep fact.
Key takeaways
- “Make it better” hands the model a judgment you have not made, so you get its average idea of better.
- The cost is rework and confused readers, not tokens.
- Before asking for a rewrite, write one sentence about what is wrong. That sentence is most of the brief.
- A good brief names the task, the reader and goal, the material, what stays and goes, the shape, the limits, an example of good, and what to do when unsure.
- Turn briefs you reuse into templates, test them on real drafts, and ask to see the prompt behind AI-drafted work.