Every previous lesson in this level starts from the same kind of request: asking the AI to write something from scratch. This one is different, and that is why it earns its own lesson: here the text already exists, you wrote it yourself, and what you are asking the model to do is read it — to tell you what reads badly, what is redundant, what is missing and where it sounds off. It is a request with its own rules, because the risk of it inventing something moves elsewhere: it is no longer in what it writes, it is in what it makes you change in your own text without needing to — a change you make on its say-so, not because it was actually wrong.
The whole thing in one sentence. When it reviews, the model has no reference
document to check your text against — it does not know whether your writing is correct, only
whether it sounds right — so its opinion on style is worth a lot and its
opinion on the facts your text contains is worth nothing, and the two have to be treated
completely differently. Confusing the two is the most common mistake in this use, and it
usually happens because both kinds of answer arrive in the same confident tone, with nothing
in the format to distinguish a well-founded opinion from an unverified claim.
What you can ask for, and what you cannot
| What you ask | Reliability |
|---|---|
| "Does this sentence read clearly?" | High. Judging clarity and structure is exactly what a language model does best, because it is pure judgement about form, not about outside data. |
| "Is anything redundant, does any idea repeat?" | High, for the same reason: it is a question about the structure of the text itself, which it has in front of it in full, not about the world outside the text. |
| "Is this fact I put in correct?" | Low. Here it is no longer judging your text: it is answering a question of fact, which is exactly the territory of the hallucinations lesson — it needs verifying against a source, not against its opinion of what you wrote. |
| "Is anything important missing that should be there?" | Medium. It can flag obvious structural gaps — an instructions document with no contraindications section — but it cannot know whether a specific detail is missing if you never told it that detail exists. |
The most expensive trap in this use. If you paste a text with a factual
error and ask it to review it, the review can easily come back focused on style and say
nothing about the underlying error — because you did not ask about that, and a model tends to
answer the question it was asked, not the one it should have been asked. The absence of a
flag about a content error does not mean there is none: it means you did not ask about it,
and it is still your responsibility to check it elsewhere.
A trick that multiplies how useful this is: give it a concrete role
"Review this" on its own tends to produce a correct but generic answer. It works much better to ask it to review from a concrete role, because that gives it a clear standard to judge against instead of a vague notion of "well written":
- "Read this as if you were a seventy-year-old patient who doesn't know any jargon." For material a patient will read, this surfaces words that feel normal to you and are actually pharmacy jargon.
- "Read this as if you were an inspector looking for something to object to." For a sign or a public post, this role is far stricter than "check if this is fine" and surfaces claims that a normal read would miss.
- "Read this as if you had ten seconds before walking past." For anything that will be read standing up or in passing — a sign, a social media headline — this role forces judging what registers at a glance, which is exactly the real condition it will be read under.
The role does not change the underlying reliability — it still cannot verify facts, only judge form — but it changes what you get back a great deal, because it gives it a concrete standard instead of letting it pick a generic one on its own, and a concrete standard is something you can push back on if the feedback still does not feel right.
A full example, and what is wrong with it
You have written a shop-window sign advertising the medication review service, and you ask the model to review it before printing.
Your original text: "Do you take more than five medicines? Come in for a complete free review of your medication, no appointment needed, and we guarantee you'll have no dangerous interactions."
What the review says
"The message is clear and the opening question hooks well. I'd suggest shortening the last sentence so it fits a sign better: something like 'complete, free review, no appointment needed'. Also consider whether 'complete' adds anything or could be cut to save space."
And now, what is wrong with settling for just that review:
- The style review is good and worth acting on. Shortening the text for a sign is sound advice, and it is exactly the kind of improvement this use delivers best.
- And the real problem, which was not a style problem, got missed. "We guarantee you'll have no dangerous interactions" is a claim no review can honestly guarantee — it is exactly what the medication review lesson (a Level 3 lesson) explains cannot be promised without knowing the case — and on top of that it may be a problematic promise from a healthcare-services advertising standpoint.
- The model did not flag it because you did not ask it to. You asked for a style review and got one; you did not ask it to check whether the content was a safe claim to make, and that is a different question that has to be asked separately, explicitly.
- The lesson: ask for two separate reviews, not one. One about style — "does it read well, is anything redundant, does it fit the space?" — and one about content — "is there any claim here I shouldn't be making, any promise that cannot be kept?" — because they are two questions with completely different reliability, and mixing them into a single request lets the style question hog all the attention, leaving the content question skimmed over even when it was the one that actually mattered.
And if your pharmacy is not like that
If what you review is the pharmacy's social media. Always add the content
question separately from the style one, more so than in any other case: a tone mistake in a
public post is fixed with an apology; an unauthorised claim about a medicine can have real
consequences, and social media is where writing fast and only skim-reviewing happens most
easily.
If several of you review the same document in turns. Use AI for the first
style pass, before the second person sees it, instead of as a substitute for that second
human review. AI never replaces a second pair of human eyes on something important; it
speeds up the mechanical work so that second pair of eyes can focus on what actually matters.
If the text is long — an internal protocol, a multi-page guide. Ask for the
review section by section, not all at once: with a long document, the review tends to
concentrate on the first few pages and get shallower towards the end, the same way it would
for a person reading quickly.
If the document is going to carry the pharmacy's or a professional's signature.
Here the content review matters more than the style review: first of all, check that no
sentence claims something you cannot back up if somebody asks you to in writing. It is the
professional-responsibility lesson applied here — the sign-off is still yours, whoever
reviewed the draft.
If the text is for internal use, nobody outside will ever read it. Here the
bar drops considerably, and a quick style review is usually enough without needing the
separate content question — the risk of a problematic claim in a document only the team sees
is much lower than in something public.
If the text was written by several people, mixing styles. Explicitly ask it
to flag tone shifts within the same document, not just whether each sentence reads well on
its own. A text written by several hands can have every paragraph correct on its own and
still read as three different people when taken together, which is exactly the hardest thing
to spot reading it yourself, because you already know every part by heart.
When it does not work first time
I ask it to review and it rewrites the whole text without being
asked.
Be explicit about what you want back: "flag the problems, do not
rewrite the text" if what you want is the list of what is wrong so you can fix it yourself,
instead of the already-changed text. Both are useful at different moments, but it is worth
choosing which one you want each time.
The style suggestions it gives me are unconvincing, they sound too
generic.
Give it more context on who the text is for and what its goal is — the
same as in the anatomy-of-a-prompt lesson. "Review this for a shop-window sign that people
will read standing up, in five seconds" produces far more useful suggestions than "review
this text" on its own.
Can I ask it to check whether the tone is too aggressive or too
soft?
Yes, and it is one of the uses it gets right most often: tone is
exactly the kind of judgement about the form of the text a model evaluates well, because
it depends on no external fact it has to verify, only on how what you already wrote sounds.
I asked for both reviews — style and content — and it only answered
the style one.
Split them into two separate messages instead of one long request. With
both questions together, it is easy for the answer to concentrate on one and treat the
other in passing; asked separately, each one gets full attention.
Does this also work for checking spelling and grammar, not just
style?
Yes, and it is the most reliable part of this whole use — spelling and
grammar are fixed rules, not judgement calls — though it is still worth reviewing the
changes it proposes before accepting them wholesale, because some "corrections" can
actually change the meaning of an ambiguous sentence rather than just its form.
Can I ask it to compare two versions of mine and tell me which is
better?
Yes, and it works well for the same thing as the rest of this use: form,
not content. It can tell you which reads better or which is shorter; it cannot tell you
which is more correct if the difference between the two versions is a fact, not a matter of
style. Ask explicitly what they differ in — tone, length, clarity — so the comparison has a
standard rather than being a generic preference.
Before moving on
- I always separate the style question from the content question, and I know they have completely different reliability.
- I know the absence of a flag about a content error does not mean there is none.
- I explicitly ask whether I want a list of problems or the already-rewritten text.
- AI never replaces a second pair of human eyes on an important document, ever.
- I give it a concrete role to read from instead of asking for a plain, generic review.
← Revisit: patient-facing material
Next: preparing team training →
← Back to the AI School