Three Of These Seven Forbid Invention. They Are The Three That Add.

Seven prompts, posted one to a comment under a single card. Each one hands ChatGPT a specific editing job instead of the useless instruction to “write better.” That part of the card is correct and worth taking seriously: a model asked to improve a draft will mostly rearrange it, while a model asked to cut every weak qualifier will cut every weak qualifier.

But read the seven together and something sharper shows up. Three of them carry an explicit line telling the model not to make things up. The other four carry nothing of the kind. That is not sloppiness. The three with the warning are precisely the three that ask the model to put something into the draft. The four without it only take things out — and a prompt that only removes material has nothing to invent with.

Whoever wrote this card put the guard rails exactly where the cliff is. Here is what that means in practice, where the card still falls short, and what the model vendor’s own documentation says about writing rules in the negative.

Sort the seven by direction of travel

Ignore the order they were posted in and ask one question of each: after this prompt runs, does the draft contain anything that was not in it before?

Three and zero is the right answer, not a coincidence

It would be easy to read the missing warnings on prompts 1, 2, 5 and 6 as an oversight — four prompts the author forgot to protect. The opposite is true. Those four are structurally incapable of the failure the guard is written to prevent.

Deleting a filler word cannot produce a fact that was not there. Shortening a sentence cannot invent a statistic. Breaking up a monotonous rhythm cannot attribute a quote to someone who never said it. The worst a subtractive prompt can do is cut something you wanted to keep, which is a different failure with a different fix — and prompts 1 and 6 both already handle it, one by telling the model to preserve the original meaning, the other by telling it to preserve essential details and personality.

The three additive prompts are a different animal. “Replace vague claims with specifics” is an instruction to produce detail that the draft does not currently contain. The model has two ways to satisfy it: find the detail, or manufacture it. Nothing in the phrasing favours the first. “Give this a defensible perspective” invites the model to reach for supporting evidence, and a model that cannot find evidence is perfectly capable of writing a sentence shaped like evidence. “Write in my style, here are samples” is the only prompt of the seven that puts somebody else’s text inside the context window, which creates a failure mode none of the others have: lifting those sentences wholesale.

Three jobs that create material. Three guards. The card did not scatter warnings for reassurance; it placed them.

Only one of the three says what to do instead

Line up the three guards and prompt 3 is plainly the best-written sentence on the card. It does not just close a door; it opens a different one in the same breath. The model is told that the correct response to a gap is to report the gap. Prompts 4 and 7 close the door and stop.

The vendors’ own fix for “don’t copy my sentences” is not a prohibition

Prompt 7 is worth dwelling on, because the exact behaviour it forbids is documented, and the documented fix is not a ban.

Anthropic’s prompting page for Claude Fable 5.1 has a section called Quoting retrieved sources. It reports that when summarising documents, the model is more likely than its predecessor to reproduce passages of the source text without marking them as quotations — the same failure prompt 7 is guarding against, in a different task. The recommended remedy is not to add a sentence saying don’t do that. It is to add one complete worked example to the system prompt: the request, a correct response, and a short explanation of why that response is correct. The page spells out what the example should demonstrate — each source conveyed in the writer’s own indirect speech, one short marked phrase at most, every other claim reworded.

That is a meaningfully different instrument. A prohibition tells the model which output to avoid and leaves it to guess at the acceptable neighbourhood. An example puts a pin in that neighbourhood. If you run prompt 7 regularly and it keeps producing sentences you recognise from your own samples, the upgrade is not a sterner warning. It is pasting in one paragraph you wrote, one paragraph the model wrote in your voice that you were happy with, and a line saying what made the second one acceptable.

And negative rules outlive the behaviour they were written for

There is a second hazard in writing editing prompts as a stack of prohibitions, and the same documentation names it twice.

On formatting, the page notes that earlier models overused bullets and bold in chat, that many prompts still carry anti-formatting rules written to suppress that, and that the current model leans the other way — using bold less, reaching for headers and lists less readily. The instruction is blunt: if your prompt contains anti-formatting language, remove it, or replace it with a rule that says when formatting is appropriate.

On progress updates it goes further. Older models narrated constantly, so people wrote system-prompt lines telling them to hold everything back until the end. Those lines are still in the prompts. The page tells you to audit for them and — its own words — “Remove lines like that before adding anything.”

This matters for a card of seven editing prompts because four of them exist to suppress habits that the current generation of models has largely dropped. Prompt 1 hunts generic phrases and predictable openings. Prompt 2 asks for speech that does not sound machine-polished. Both were written against a real problem. If you run them against a 2026 model on a draft that is already plain, you are not removing AI tells; you are paying for a pass that finds nothing, or worse, one that strips voice the draft needed. The guards on prompts 3, 4 and 7 age well because invention is a permanent property of the machine. The de-robotising prompts age badly because the robot keeps changing.

Where the risk actually sits

Run the seven in the order the card gives them and the sequence has a clean shape. Everything subtractive could run first, everything additive second, and the single moment worth your attention is the boundary between them.

Compared with the card that had no additive prompt at all

An earlier card from the same page offered eight humanizer prompts, and every single one of the eight was subtractive — eight ways to take the robot out of a draft, not one slot where the person went back in. By that standard this card is a clear improvement. Three of its seven prompts put something in, and prompt 7 is the only prompt across either card that asks you to supply your own writing as raw material.

The trade is that the new card inherits a problem the old one could not have. Once a prompt adds, somebody has to check what was added. The instinct behind prompt 3 — report the gap instead of filling it — is the same instinct behind asking a research model to tell you what it could not find, and it is the single most transferable idea on the card. It belongs in prompt 4 too.

Four honest caveats

1. The sort is mine, not the card’s. The card presents seven prompts as a flat list and never labels any of them additive or subtractive. The 3-and-4 split is a reading of what each prompt does to a draft, and prompt 2 is genuinely arguable — rewriting for natural speech reshapes sentences rather than purely removing them. I have counted it as subtractive because it operates on existing material; you could reasonably count it the other way, which would make the pattern four and three rather than three and four.
2. I have not tested the prompts. Everything above is an analysis of the wording. Whether prompt 3 actually produces flagged gaps rather than invented specifics on a given model, at a given temperature, on your draft, is an empirical question I have not run. Prompt phrasing is weak evidence of prompt behaviour.
3. The documentation I have quoted is Anthropic’s, and the card is about ChatGPT. The Quoting retrieved sources and formatting passages are from the Claude Fable 5.1 prompting page, verified on the live page. They describe a documented failure pattern and a documented remedy, and the reasoning generalises, but they are not statements about OpenAI’s models and I am not presenting them as such.
4. Nobody has verified the prompts’ provenance, including me. The card does not say where these seven came from or what they were tested against. Treat them as a well-constructed starting point from an anonymous source, which is what they are.

The one-line version

Four of these prompts take things out of your draft and three put things in, and the three that put things in are the three carrying a warning about invention — so run the four first, read what is left, and give the last three the attention the card implies they deserve.