Native learning from your edits

By Jack Stovell X · Instagram · LinkedIn · published 2026-09-20 · updated 2026-09-30

You edit an AI draft the same way every time. You shorten the sentences. You put the contractions back. You cut "furthermore" and the em dash and the third item in the list. You add a word the model would never choose. Then you send it, and the next draft arrives with all of the same problems, because the tool never saw what you changed.

ScriptGrain does. Every draft you finalise is compared with what was generated, and the difference is measured: which habits moved, which words you added, which you removed every time they appeared. Those measurements become proposed changes to your profile, in an inbox, with the evidence beside each one. Accept them and the next draft starts closer.

See the loop on your own writing. Paste an AI draft and the version you actually sent. The difference is measured in your browser and turned into the same suggestion shapes the product proposes. Nothing is uploaded.

What an edit can teach

An attribute to adjust. Your edit moved average sentence length from 24 words to 14: the profile's target was wrong, or the draft ignored it. Either way the number should move.
A word to prefer. You added "properly" three times to a draft that never used it. That is vocabulary the profile did not know was yours.
A word to drop. You removed every "leverage". It goes on the banned list, which overrides everything else in generation.
A rule. You cut all eleven em dashes. "Never use em dashes" is now a hard rule for this profile, not a preference.

What it deliberately does not learn

A light edit teaches nothing, on purpose. Suggestions carry a confidence floor and need several edits pointing the same way before they reach you, so one unusual piece does not bend the profile. And nothing changes until you accept it: the inbox is a proposal, the profile is yours.

Why "native" matters

Most tools let you edit the output and forget it. Some let you thumb a draft up or down, which tells the model nothing about what was wrong. Reading the edit itself is the signal: it is the most specific feedback a writer ever gives, and they give it on every piece without being asked.

What a saved edit looks like to the profile

The profile never sees your editing session. It sees two documents. The draft as generated, and the version you saved as final. Everything it learns comes from the gap between them.

So the final version matters more than you'd think. If you tidy a draft in one window, then paste a slightly different version into the tool before sending, the profile compares against the wrong thing. Save what you actually sent. The file that left your hands is the evidence.

Picture a draft with a dozen long sentences and two "on top of that"s. You split the sentences, cut both, and add "properly" twice. Saved, that becomes a measurable difference: sentence length moved, two words removed, one added. Nothing more mysterious than that.

And to be fair, the profile can only read what's on the page. It can't see that you were tired, or that you hate a word for reasons of your own. It sees that the word was there and then it wasn't. Clean evidence in, useful proposals out.

Edits that teach and edits that do not

An edit teaches when it points the same way again and again. Cut "leverage" every time it appears, across several pieces, and that's a pattern. Put the contractions back in every draft, and that's a pattern too.

An edit does not teach when it's a one-off. Say you write a note for a formal audience and stiffen the whole thing: no contractions, longer sentences, a more careful tone. That's a fair choice for that reader. It's also not your voice. If you save that version as your final, you're telling the profile your voice changed.

Same goes for a rewrite driven by the topic. A legal explainer needs terms you'd never use in a newsletter. Those words are the subject's, not yours.

So the useful test is simple:

Would I make this change in most things I write?
Am I fixing the draft, or fitting one particular reader?
Have I made this same change before?

Yes to the first and third, go ahead and save it. Yes to the second, think twice.

How to edit so the profile learns faster

Be consistent. That's most of it. The profile looks for edits pointing the same way, so the fastest route is to fix the same problems the same way each time. If you shorten sentences on Monday and lengthen them on Thursday, the evidence cancels out.

Save the final version every time, even when the draft was close. A light edit teaches little, but it costs you nothing, and the pieces add up.

Keep audience-specific rewrites out of the loop. If you've adapted a piece for one client or one occasion, finish and send it, but don't treat it as a voice example. You can still use it; it just isn't evidence of how you write by default.

Then use the inbox properly. Each proposal arrives with its evidence beside it. Read it. If "never use em dashes" matches what you do, accept it. If a word was cut only because it clashed with one headline, decline it. Nothing changes until you say so, and the profile is better for the judgement you put in.

Edit the way you'd want the next draft to read. Then let the pattern build.

Questions

Do I have to do anything for the profile to learn?

Finalise the draft in ScriptGrain (or send it back through the API's save_edit). The comparison and the proposals happen on their own; you accept or dismiss them on the profile page.

Can a bad edit make the profile worse?

One edit cannot: suggestions need repeated evidence and clear a confidence floor before they are shown, and none apply without your acceptance. Version history lets you restore an earlier profile if you ever change your mind.

Does this work with drafts written in ChatGPT or Claude?

Yes, through the connector or the API: a draft generated or scored through ScriptGrain, then saved back with your final text, feeds the same loop.