The Edits Are the Training Data

I have been thinking about what actually happens when I keep rewriting the things AI gives me for Raw Drafts. The obvious answer is that the article gets better. The first version is usually too clean, too generic, too certain about what it is trying to say. I remove parts, add examples, make the paragraphs less mechanical, and sometimes change the direction of the whole piece. Eventually it begins to sound more like something I would have written.

But the article is not the only thing being improved.

Every change I make also tells the system something about what I think quality is. The final version shows what I kept, but the edits show what I distrusted. They reveal which phrases sounded false, where the argument became too neat, where the writing lost contact with something real, and which kinds of conclusions I keep removing because they arrive before I do.

That may be more useful than the final article itself.

It is easy to tell an AI to use my voice, but that instruction does not mean very much on its own. A voice is not a font or a set of words. It appears in the choices made while thinking. It is visible in how long I stay with an idea, where I become suspicious of my own argument, what kinds of examples I trust, and what I decide is still unresolved.

Those things are difficult to describe in advance.

I can say that I want full paragraphs, that I do not like polished little punchlines stacked on top of one another, and that I do not want every article to end with a clean answer. These instructions help, but they are still general. The actual edit is much more specific. It shows the exact sentence I removed and what I replaced it with. It shows that this claim was too broad, that this paragraph was pretending to know more than it did, or that the writing became more interesting only after I added something that actually happened.

The edit contains the judgment.

Most writing systems do not preserve much of this. We keep the final version and discard the route that led there. A paragraph may be rewritten fifteen times, but the reader only sees the last one. The weak versions disappear, along with the reasons they were weak. The result survives. The process becomes invisible.

Software has treated this differently for a long time. Code usually has a history. You can see which lines changed, when they changed, and sometimes why. The current version is only one point in a sequence of decisions. Writing has mostly treated drafts as temporary material, but AI makes that seem increasingly wasteful because those drafts and corrections can influence the next thing the system produces.

Each correction is a small piece of preference data. This sounds like LinkedIn. This is too abstract. This example is better than the explanation around it. This ending makes the article sound more resolved than I feel. This sentence is technically fine, but it does not belong in this piece.

None of these comments looks important alone. Together, they start to describe a way of thinking.

A model trained only on the published Raw Drafts articles might learn some visible patterns. It could imitate the sentence length, recurring themes, and some of the rhythm. It might produce something that looks close from a distance. What it would not understand is the boundary around the writing. It would not know what almost made it into the article and why it was removed.

That boundary is probably where much of the voice lives.

I am more interested in an AI that learns from everything I rejected than one that only learns from everything I published. The rejected material shows where plausible writing failed. It shows the moment a clever sentence became distracting, or when a strong conclusion flattened an idea that was still open. It captures the resistance that shaped the final piece.

This makes Raw Drafts feel like it could become something slightly different from a normal publication. The articles would still matter, but underneath them there could be an editorial memory. Drafts, changes, comments, repeated corrections, and patterns that become visible only after many pieces. The site would not only contain what I thought. It would also contain how I decided what was worth keeping.

I do not think this should become a rigid style guide. That would probably make the writing worse. A lot of my decisions depend on the context of the article. A sentence that feels too certain in one piece may be exactly right in another. Sometimes structure helps the thinking, and sometimes it hides the fact that the thinking is weak. The system does not need to turn every edit into a universal rule. It may only need to remember them well enough to become less surprised by what I change.

That already feels useful.

The first draft is becoming cheap. It takes almost no effort to generate another version. The difficult part is still recognizing what is wrong with it. The value may be shifting away from producing language and toward making better decisions about language.

Editing used to feel like the work done after writing. Increasingly, it feels like the work itself.

A good edit improves one article. A history of good edits may improve every article after it. That history becomes a kind of accumulated taste, built from hundreds of small disagreements with the machine.

Maybe this is how personal AI systems will actually become personal. People will not sit down and formally train them. They will work normally. They will rewrite, reject, shorten, expand, question, and move things around. The useful data will come from all the moments where the system was almost right.

The draft proposes something. The edit shows what I prefer. Over time, the pattern of those edits begins to describe me more accurately than the final articles do.

So every time I change a Raw Drafts piece, I may be producing two things. I am improving the article in front of me, and I am leaving behind a record that could improve the next one.

The published article is the visible result.

The longer-term work may be teaching the system how I decide.

Connected reading

• Raw Drafts Are a Form of Learning
• The Real Test of AI Code Is the Next Prompt
• AI Slop Is the Factory. Brainrot Is the Product.
• The Backlog of Possible Lives

AI authorship note

This essay began with Kris Haamer noticing that the changes he makes to AI drafts are themselves a record of how quality improves. AI helped develop the idea into an article and repeatedly produced versions that were too segmented, polished and unlike his established Raw Drafts voice. Kris pushed the writing toward longer connected paragraphs, changed the title to “The Edits Are the Training Data,” and approved the final publication.