What a good AI-written first draft should actually look like

Ask most AI writing tools for a LinkedIn post draft and they hand back something polished. The sentences are clean. The transitions work. The ending lands. It reads as publishable. Many users do publish it, with minor edits.

This is the wrong target. A draft that arrives polished is a draft that cannot be edited, because the writer has no clear handles to grip. The draft is a finished object that is simply not the writer's object. It requires replacement, not refinement.

A good first draft looks almost the opposite. It is visibly rough, with the roughness pointing in a direction. The writer sees what is missing, what needs sharpening, where to add a specific example. The editing handles are obvious because the draft is an invitation to edit, not a fait accompli.

粗Rough is a feature, not a bug.

A finished-looking draft implies a decision has been made. This is where the piece opens; this is what the ending is; this is how the argument goes. The writer reading the finished draft has a binary choice: accept these decisions, or start over. Accepting them means publishing a piece shaped by the tool's defaults. Starting over means writing from scratch after paying the time cost of reading the tool's output first.

A rough draft implies the decisions are still the writer's to make. The opener is a placeholder. The ending is tentative. The middle could go either way. The writer's job is to sharpen: pick the right opener from three rough options the draft gestures toward, decide which of two endings fits, add the specific example that grounds the middle section.

This is not the draft being worse. It is the draft being the right kind of unfinished. Rough drafts invite editing. Polished drafts invite replacement.

The craft has always known this. A good editor, handed a rough draft from a writer they know, sees where the gaps are and fills them from their sense of what the writer would want. A good editor handed a polished draft has no gaps to fill; they can only rewrite. Editing is downstream of the draft's roughness, not its smoothness.

形In the writer's shape.

The other feature a good first draft must have is that the roughness points in the writer's direction, not in some neutral AI direction.

This is the feature that distinguishes a good AI draft from a bad one. Both are rough. The bad one is rough in no specific direction. The ending placeholder is a generic "In conclusion..." The opener placeholder is "Here's what I've been thinking." The argumentative move in the middle is "On the one hand... on the other hand..." These defaults are directionless; the writer editing them is choosing from infinite possibility because no specific shape has been gestured at.

A good draft is rough in specific ways that reflect how the writer actually writes. The ending placeholder is a short declarative observation, because the writer closes on short declaratives. The opener is a specific condition the reader is in, because the writer opens that way. The argumentative move is a concede-and-pivot, because the writer prefers concede-and-pivot to claim-counterclaim.

The roughness still gives the writer work to do. What the roughness also provides is a shape to work within. The writer is not choosing among infinite possibilities; they are sharpening a draft that already knows how they typically write and is handing them the first approximation of that shape. The editing is the writer's own habits, applied to their own draft.

時The 15-minute edit target.

The target for the editing pass, on a good first draft, is roughly fifteen minutes for a 250-word LinkedIn post. This number is worth being specific about because it has consequences for whether the whole workflow makes sense.

Forty-five minutes is the cost of writing from scratch. Twenty minutes is the cost of editing a generic AI draft, which ends up being rewriting in practice. Fifteen minutes is the cost of editing a voice-matched rough draft, where the work is surface tightening rather than structural replacement.

The ten-minute swing between twenty and fifteen minutes matters. Multiplied across weekly posting over a year, it is the difference between publishing fifty-two posts and publishing twenty-five. It is also the difference between the posts reading as the writer's and reading as the tool's. The writer's editing time is where the writer's voice enters; compressing the writer's editing time into a smaller share of a longer rewriting process (the 20-minute workflow) means less of the writer's voice gets into the final output.

Fifteen minutes is long enough for the writer to do real work. Pick the specific example. Sharpen the claim. Tighten the close. Remove the one word that does not fit. Add the aside that the draft gestured toward but did not complete. This is the work editing was designed for. The writer comes out with a post that is provably theirs, in a time budget a busy professional can actually afford.

要What a tool has to know to produce this.

Producing a draft that is rough in the writer's direction, rather than rough in no direction, requires the tool to have a model of how the writer actually writes. A generic model does not have this. A prompt ("write in my voice") does not give it to the model, because the writer's voice lives in the structural habits across hundreds of pages of their writing, not in a two-paragraph description.

The tool needs three things.

First, a persistent profile of the writer's structural habits, built from a substantial corpus of the writer's actual work. Not a self-description; actual writing. The profile captures opening patterns, argumentative moves, closing shapes, and source-integration tendencies across enough samples to generalize.

Second, a generation process that uses the profile as constraints during drafting, not as a post-hoc polish. A model that writes the draft and then tries to "voice-adjust" it produces generic output with voice decals on top. A model that writes inside the voice-profile constraints from the first token produces drafts with the writer's structural signature embedded.

Third, a feedback loop that updates the profile over time based on the writer's edits. If the writer consistently sharpens a particular kind of closer, the profile should learn that this is the writer's closing preference. If the writer consistently softens a particular kind of claim, the profile should learn the writer's hedge pattern. The profile should get more accurate with each edit, rather than staying static.

This is the set of capabilities Onaji's Voice Profile is built around. The draft is rough; the roughness is in the writer's direction; the fifteen-minute edit passes are surface tightening, not structural replacement. The feedback loop refines the profile every week.

The practical consequence.

A writer who adopts a tool built around these principles should expect the drafts to feel different from what generic AI tools produce. The drafts will look less finished. The drafts will also feel more tractable. The editing pass will not be a fight.

Over months, the writer's published posts will read as theirs. Their feed will not drift toward the tool's voice, because the tool was never imposing a voice to begin with. The tool was amplifying the writer's own, applied more consistently than the writer could apply it themselves every Tuesday morning at six.

The test of a good first draft is not whether it looks publishable. It is whether the writer, ten minutes in, is improving their own thing rather than fighting someone else's.