We're Saving the Wrong Thing
An AI draft is worth everything for about twenty minutes, and nothing after — yet the draft is the thing we keep, while what it revealed about your intent evaporates with the session. An essay on the Delta Principle: drafts are ephemeral; deltas accumulate.
A few months ago I caught myself doing something strange.
It was late. I had asked Claude for a product brief, and an 1,800-word draft came back — almost right, the way AI drafts usually are. Five things were wrong: a bad assumption in section two, a missing edge case, a tone problem in the third bullet, a dependency that doesn't exist, a rollout step out of order. I marked all five, sent them back, and a new draft arrived.
Here's the strange part: I never looked at the old draft again. Not once. A document I had just read with real care — more care than I give most emails from actual humans — was worthless twenty minutes later. Nobody archived it. Nobody mourned it. And this wasn't an exception. It happens to me a dozen times a day now, and if you work with AI, it happens to you too.
We don't have a name for what's going on in that moment. So let me just say what I think it is:
We're saving the wrong thing. Fifty years of knowledge work trained us to treat the document as the asset — write it, version it, archive it, measure progress by it. But in a growing class of AI-native work, the draft is not the durable unit. The reusable unit is the delta it helps reveal — and that's the thing we let evaporate while we dutifully keep the drafts.
The rest of this essay is that claim, unpacked: where it holds, where it breaks, and a way to test it.
What actually survived
Look at what actually happened in that late-night session. That draft's job was never to be kept. Its job was to be judged. The moment I finished marking it up, everything valuable in it had been extracted — and what was extracted wasn't text. It was what the model learned about me:
- The audience is product managers, not engineers.
- Be direct. Drop the marketing tone.
- That dependency doesn't exist. Stop citing it.
- The rollout order is not negotiable.
Notice that none of these are edits. "This sentence was deleted" is an edit — it tells you what changed, not why, and it tells you nothing about the next draft. "The audience is product managers" is different in kind: it's a correction to the model's picture of my intent, and it applies to every draft that follows.
That's the unit that survives the draft. It deserves a real definition, so here is one:
In practice, a good delta has two properties. It's scoped — "prefer direct language" is true for this brief, not for every word you'll ever write, and a delta that doesn't know its own boundaries misfires outside them. And it's anchored — tied to the exact passage that provoked it, because "keep the tone of this paragraph" means nothing without the paragraph. Feedback that loses its anchor decays into a fortune cookie. Which is why a delta carries its anchor with it: the quoted passage travels inside the entry, so the log stays meaningful long after the draft it points into is gone.
A single critique produces a handful of deltas. The accumulated record of them is a delta log — and the log, not any draft, is the asset worth keeping. A delta is one entry; the log is the wealth.
A diff stores a change to text. A delta stores a change to intent.
The database people saw a version of this
The analogy is useful but imperfect, so I'll keep it short.
Before databases, you saved files; the file was the asset. Then transaction logs arrived, and in a whole family of systems — write-ahead logs with their checkpoints, replication, event sourcing — something quietly flipped: the durable thing became the log, and the table you look at became a view, rebuildable from the log plus a checkpoint. Martin Kleppmann calls this "turning the database inside out." Ask an engineer who runs one of these systems what's sacred, and they'll say the log, not the table.
Something similar is happening to working drafts: the delta log becomes the asset; the draft becomes a view. In this kind of workflow, a draft is not the work itself. It is a projection of the model's current picture of your intent.
Here's where the analogy stops, though — and the gap is instructive. A transaction log is complete and deterministic: replay it and you get the database back, bit for bit. A delta log is deliberately partial and makes no replay guarantee. You don't record everything that happened in a review — you record what generalizes. In that respect it's less like a transaction log and more like learning itself: a system that learns doesn't remember every example it saw. It keeps what the examples taught it.
The Delta Principle
So here is the principle:
In this class of work, an intermediate draft is not a knowledge artifact. It is an elicitation surface — a probe whose purpose is to extract judgment from you. That's a quiet inversion of what documents used to be: in the old world, you wrote a document to communicate information to a reader. Many working drafts in an AI workflow have no future reader at all. They exist so that a human will react to them, and the reaction — not the draft — is what feeds the next iteration. The draft communicates in order to extract.
Working with a model is a long game of twenty questions. Each draft is a question. Your feedback is the answer, and every answer narrows what the model believes you actually want. Progress isn't the version number going up; it's the space of misunderstanding going down. (The machine-learning crowd will recognize this as active learning wearing a trench coat — except the query the model poses is an entire document.)
Why drafts still have to be good
An obvious objection: if a draft is just a question, why should it be any good? Why not generate garbage and harvest the feedback?
Because a bad probe extracts noise. Hand someone a terrible draft and the delta you get back is "start over" — which contains almost nothing about intent. Hand them a perfect draft and you get no delta at all; there was nothing left to learn. The most informative draft is the near miss: good enough that your feedback is about intent rather than competence, wrong enough that there's something real to extract.
So the principle doesn't excuse sloppy drafts — it explains what a good draft is for. The iteration value of an intermediate draft often lies less in its standalone quality than in the quality of judgment it elicits. And that hints at a research question someone should chase: what if you optimized a model not for the quality of its first draft, but for how much it honestly learns per round of human feedback?
The Regeneration Criterion
A review distills an 1,800-word draft into a handful of deltas — a few dozen words each, generalizing far beyond the draft that produced them. Which raises the obvious question: how do you know the deltas you kept are enough? That question has a checkable answer, and it's what makes this a framework rather than a mood.
Take a document you iterated on with an AI. Gather the source material and the delta log — and none of the intermediate drafts. Hand them to a fresh model, one that has never seen the conversation, and see what it converges on.
But don't stop at "did it produce something acceptable" — a strong model might manage that from the source alone, and then the log proved nothing. The real test is comparative, at an equal token budget: does source + delta log beat source-only? Does it beat a generic summary of the conversation? And against the full draft history, the question isn't whether the log wins — more information usually wins — it's whether a log a tenth the size gets you nearly all the way there.
A delta log earns its keep when a fraction of the tokens buys almost all of the outcome. That's the criterion. It's checkable, and it means I can be proven wrong — which is what separates a framework from a slogan.
Where this breaks
A claim is only as credible as its boundaries, so here are the places this one doesn't hold:
- Creative work where form is the essence. A poem's fourth draft isn't a probe; the exact words are the point. You can't reduce craft to intent.
- Legal and contractual text, where the precise wording is the asset and every historical phrasing may matter later.
- Science and audit trails, where full provenance — who saw what, when — is the requirement, and partial records are disqualifying.
- Design artifacts that act as boundary objects — a mockup a team gathers around isn't just eliciting intent; it's coordinating people. Throwing it away throws away the coordination.
- Multi-stakeholder work with conflicting intents, where the log stops being one person's judgment and becomes a negotiation record — a different, harder object.
- Work where the act of creating changes the creator. Philosophy, fiction, research: writing the draft doesn't just reveal what you intended — it produces intentions that didn't exist before you wrote it. A log can record where you landed; it can't replace the walk.
If your artifact is the deliverable, keep the artifact. The claim here is about the drafts in between — the ones nobody was ever going to read again — and about what we should be keeping instead. Every document is ephemeral except the last one; this essay is about all the others.
From artifact-centric to learning-centric
Inside those boundaries, a real shift is underway: this class of knowledge work is moving from artifact-centric to learning-centric. The artifact is no longer the center of the process; the learning is. And the uncomfortable part is that today, the learning has almost nowhere good to live. The weights don't update. The context evaporates. Yes — the major assistants now ship memory features, and they will dutifully record that you prefer direct prose. But hold that memory up against the definition above: it's unscoped (applied to everything you'll ever write), unanchored (detached from whatever provoked it), unauditable, and locked inside one vendor. It's exhaust with a UI. A delta that exists only as vibes in a chat history — or as a one-line preference in a platform's memory pane — isn't an entry in a log. For intent to compound, deltas have to be explicit artifacts: scoped, anchored, written down, portable across sessions and across models.
And none of this is only about prose. Code you review, designs you critique, strategy decks, UI mockups — wherever you iterate with an AI on an artifact nobody will read twice, the same loop runs: the artifact is the question; the judgment is what's worth keeping.
A lens, not an invention
I should be clear about what's being claimed. None of the ingredients here are new. Critique is old. Iteration is old. Preference learning, feedback, memory — all old, all studied. What I'm offering is a lens: one principle that ties familiar phenomena together — drafts as elicitation surfaces, judgment as the durable unit — and one practical test, the Regeneration Criterion, that tells you whether your way of working actually captures what mattered. If the lens makes a dozen scattered habits suddenly look like one thing, it has done its job.
Where this ended up
This lens is the design principle behind a tool I built, called PassbackAI — one attempt to treat deltas as first-class objects: anchored comments that travel back to the model as structured instructions, rather than drafts piling up as dead versions. I used to describe the absence of an archive apologetically. I've stopped.
To be clear about the boundary: the tool closes a single loop today — one draft, one set of anchored deltas, one pass back. The cross-session, cross-model delta log this essay argues for is the part nobody has shipped yet, including me. That's not modesty; it's the roadmap.
Surfaces and assets
In AI-mediated work, many drafts are not things we mean to preserve. They are surfaces we use to reveal intent. The mistake — the one we're all making by default, a dozen times a day — is treating every surface as an asset while the intent it revealed evaporates with the session.
The artifact may be temporary. The intent it reveals is what compounds.
We already know how to archive documents; we've had fifty years of practice. The next generation of AI tools will be judged by something else: how well they keep what the documents taught us.
As for that late-night brief — I still can't tell you what its first draft said. I can tell you exactly what it taught the model. That was the part worth keeping.
Postscript: this essay was written the way it says work now happens — drafted by a model, steered by round after round of my anchored feedback. If you asked me what the third draft said, I'd have to go dig. The deltas I can recite from memory. That was the point.