"Generate 30 posts about productivity" and "turn this article into five posts" sound like variations of the same request. They are not. One asks a model to invent judgement; the other asks it to find the judgement that already exists in your writing. The second is almost always the better use of the tool, and the difference explains why so much AI-assisted content reads as plausible and weightless.
Generation invents a position
When you ask for content about a topic, you get the average of everything written about that topic. That average is fluent, well-organised and interchangeable — it is the reason readers increasingly skim text that sounds like it, without being able to say why.
The failure is structural, not a matter of prompt quality. A model has no access to the specific thing that makes your writing worth reading: the decision you made, the number you measured, the option you rejected. Ask it to generate and it must fill that gap with something generic.
Repurposing extracts a position you already took
Repurposing starts from a different premise: the thinking exists, and the job is to find where it is publishable.
That changes what the model is doing. Instead of inventing a claim, it locates the claims in your draft. Instead of inventing evidence, it keeps the numbers attached to the sentences they support. Instead of inventing a voice, it preserves the identifiers, the named failures and the constraints.
The output is still written by a machine, but every substantive claim was already true before the machine touched it. That is the property readers are actually detecting when they say a post "feels real".
A practical division of labour
| Task | Better suited to |
|---|---|
| Finding the five angles inside a 2,000-word draft | Machine |
| Deciding which angle is worth publishing first | Human |
| Rewriting a Chinese paragraph into English sentence order | Machine, with a human check on the first two lines |
| Choosing which customer story may be told publicly | Human |
| Drafting a platform-native first line from a known claim | Machine |
| Knowing whether the claim is defensible | Human |
The pattern: the machine is good at search, restructuring and phrasing. It is bad at knowing what is true about your business, and it has no way to acquire that.
Three failure modes to watch
Smoothing. Models remove specifics to make sentences flow. "We lost the account in week three" becomes "we experienced churn challenges". Guard against it by keeping numbers and named events in the prompt and checking that they survive the output.
Hedging. Generated text adds qualifiers until claims become inoffensive. In English, hedged claims read as "the author is unsure", which is the opposite of what a founder wants.
Volume without inventory. Generation makes it cheap to produce thirty posts, which is exactly why it tempts you to publish below your own bar. A queue of extracted units keeps the bar where it was.
What this means for pricing your time
If your only goal is filling a calendar, generation is cheaper. If your goal is being the person whose posts get forwarded, repurposing wins, because forwarding depends on specificity and specificity cannot be generated.
A useful test before publishing anything AI-assisted: point at the sentence a reader would quote back to you. If you cannot find one, the piece was generated rather than extracted.
Three rules that keep the judgement yours
- The sentence has to survive without the model. Before publishing anything AI-assisted, point at the line a reader would quote back to you. If you cannot find one, the piece was generated rather than extracted — and a page saying nothing specific is exactly what search engines and readers both skip.
- Pin the numbers and the named events in the prompt. Smoothing is the failure mode: "we lost the account in week three" quietly becomes "we experienced churn challenges". Keep the specifics in the instruction, check them in the output, and the post keeps both its credibility and its search value.
- Disclose what a reader would feel misled by. Using a tool to restructure an argument you actually made is not the same as publishing a position you never held. The first is editing; the second is fabrication, and it costs more than the reach it buys.
- Check two places by hand, not twenty. The first two lines and the numbers — those are where a translation, a hedge or an invented figure does the most damage, and both take under a minute to verify.