A lot is moving at once.

Anthropic shipped Opus 5.5 on Tuesday, and OpenAI shipped GPT-6 Sol and GPT-6 Luna an hour later.

The week before, TypeSafe shipped Jev, a model that answers with a typed decision and a probability attached instead of a paragraph. I wrote about Jev and that probability on Monday.

Any of these may get an issue of its own. This one is about something older than all of them.

Your prompt document is on version seven, and somewhere in its filename is the word FINAL, possibly twice.

Each version in that file is a correction you made because a real draft came back wrong in a real way.

And the correction still did not hold.

Two jobs, one home

A prompt is how you ask, and after seven rounds of sharpening, yours is probably a very good ask.

But asking and checking are separate jobs. An instruction can be perfectly clear and still never get compared against the draft that comes back.

I find it easiest to see in a table. Here is where each job lives in a setup built on a prompt document.

The job	Where it lives	What happens
Asking well	The prompt document	Works, every time you remember to paste it
Checking the next draft against what you decided	Nowhere	Decided again, from scratch, in the next chat
Refusing a draft that broke the rule	Nowhere	You catch it, or nobody does

Look at the second row. The answer is not "done badly." It is nowhere.

No sentence you could add in version eight gives that job a place to live.

Why the sharper prompt kept looking like the answer

A sharper instruction holds for a few drafts, which is exactly long enough to be convincing.

You write it, the next drafts come back right, and the wording looks solved.

Then the context window turns over, or you start a new chat, or the tool updates, and the next draft arrives without it.

So you sharpen again, and version eight joins the file.

Every pass through that loop improved the ask. Nothing in the loop ever checked the answer.

What checking looks like

The fix is a different kind of object, not a longer document.

It is a rule a draft can fail, written once in a file you can open, with the reason attached, so that six weeks from now you can still tell whether it applies.

A description only matters when someone remembers to paste it. A rule written where a draft can break against it is already there when the next draft arrives.

That is the smallest specific version of a business that learns: a correction you own, written where the next draft can break against it.

Version seven is a very careful description.

Try this

Name the last thing you re-taught your AI.

Not the category. The sentence. "Stop opening with 'I hope this finds you well.'" "Never call the client a partner." "Use this year's figures, not last year's."

Notice how fast it came to you, word for word.

Then look at where that sentence lives right now, and ask what would compare the next draft against it.

Send me the sentence

Hit reply and paste it. Just the sentence, the way you would say it to the tool.

I read every one.

I am collecting them because the ones that show up across completely different businesses are the ones worth writing as a rule a draft can be checked against, instead of a line you retype.

Excelsior,

Pierre/
Founder, Curio Chat Academy

P.S.: Read version one and version seven side by side. The differences are a record of your standards, written in your own words, and nothing checks a draft against any of them.

P.P.S.: If the place this keeps happening is your inbox rather than your drafts, that is a different job, and I run it as a done-with-you engagement on your machine with your credential. Details: Inbox Control.