Stop asking AI to improve your writing
Asking for improvement and asking for an audit are different tasks. The workflow I use to find problems without outsourcing the editorial decision.
Asking for improvement and asking for an audit are different tasks. The workflow I use to find problems without outsourcing the editorial decision.
When you reread your own text, there is a problem that is hard to notice: you know what you meant to say.
Your brain completes the sentence before it finds the flaw.
My old reflex was to ask an AI to “improve the text.” Today I try to ask for something else.
When you ask for improvement, the model returns another version.
It may be clearer. It may be shorter. It may even be better.
But there is a side effect: the editorial decision starts to shift from you to the model.
When I ask for an audit, the work is still mine.
I want the AI to point out the problem and stop there.
These are the categories I usually look for:
| # | what to look for | how it shows up |
|---|---|---|
| 01 | unsupported claim | a number, data point or factual claim with no source or context |
| 02 | internal contradiction | two passages that seem to argue for incompatible things |
| 03 | open promise | the title promises something the body does not deliver |
| 04 | unstable term | the same word changes meaning over the course of the text |
| 05 | stalled paragraph | the text repeats the previous idea without moving forward |
| 06 | generic passage | it could appear in almost any article on the subject |
| 07 | late answer | the main question is only answered after a long introduction |
The goal is not to have the AI “grade” the text.
It is to turn it into a second pair of eyes.
My base prompt starts like this:
Do not rewrite anything.
Quote the exact passage before commenting.
At the end, list only the 3 most serious problems.
The third instruction matters.
A list of twenty problems looks thorough, but it also makes it more likely that you will accept everything on autopilot or ignore everything.
With three, you have to make a judgment.
AI can point out that a number has no source. That does not mean the number is wrong.
It can find an apparent contradiction that makes sense once you know the context.
It can call a passage generic because it has no access to the reason you chose that example.
That is why the audit is for finding places to look.
Checking, deciding and publishing are still human work.
New names for AI optimization help organize the conversation when they come with a scope. Three questions to assess any recommendation: which system, which information and which signal to track.
With AI, a content idea can be fully written before anyone decides whether it deserves to exist. Criteria for telling search intent apart from keyword variation, and a reminder that every publication requires maintenance.
A simple Search Console filter to find queries that already bring visibility but do not yet turn that visibility into clicks.