Essay · How I work
The first draft always reads finished
The three-pass review I run on every AI draft, and what it catches.
At WB Games Boston I wrote launch email for Lord of the Rings Online and Dungeons & Dragons Online. A big part of the job was catching copy that promised more than the game could deliver. A sentence would pass through enough hands in marketing and production that a plain product description turned into a promise we couldn't keep.
AI drafts have the same problem, only faster.
Where AI drafts break
A bad forgery gets stopped at the counter. A good one circulates for years, until someone finally checks it, and whoever is holding it pays.
AI drafts are the good kind. They fail by sounding right, and you find out after you hit publish.
A model will happily write “cuts research time per prospect from 20 minutes to 2” because that sentence has the shape of a claim buyers respond to. Shape is what AI is good at. The same shape fits a true number and a made-up one equally well.
The EBU and the BBC had journalists at 22 public broadcasters grade more than 3,000 AI answers about the news. 45% had at least one significant issue. 20% had major accuracy problems, including made-up or outdated details. (Source ) And the model had the article sitting right there.
Why the model can't check itself
Ask a model to check its own claims and it'll tell you they're fine. Researchers tested exactly that.
Language models “struggle to self-correct their responses without external feedback, and at times, their performance even degrades after self-correction.”
The model has no way to know whether a claim is true. Grading its own work just gets you a second confident answer. Somebody has to walk in with the source document and the pricing page, and with a real sense of who's reading.
Before I prompt
Most of the work happens before the draft. Who is this for? What does it have to do? I pull the sources and go looking for the product's limits, usually features locked behind a higher tier or places where the product is weak. Then I tell the model to flag what's missing instead of filling the gap.
That helps. It doesn't fix it. A model told to flag gaps will still write a confident sentence, then another one explaining why the first one matters.
The review pass
- Pass 1 · Check the evidence
I open the sources and check every claim. A number like 20 minutes to 2 usually traces back to one customer on one call. So I narrow the claim to what the evidence supports, or I cut it.
- Pass 2 · Check the promise
The promise has to describe what the product actually does, however much feeling it carries.
- Pass 3 · Check what's missing, then the prose
Does the reader have everything they need? Then rhythm and specificity. If a competitor could paste the same sentence onto their own site, it goes.
Claims come first because pulling one usually changes everything around it.
Hedging fails too, just more quietly. “Can help reduce costs” can't be proven wrong, so it isn't really a claim, and readers treat it that way. If the evidence supports “does,” I write “does.” The sentence gets shorter, and now it's something a customer can hold you to. Deliver on it and you earn the repeat customer.
Last, I read the whole thing once more, headline included. Headlines are where claims get squeezed hardest and checked least. This read is more feel than logic. Copy works a lot like music. You learn the rules so you know when to break them.
What it costs
The pass is slow. On a long piece it can take longer than the draft did, because checking a source takes longer than reading a sentence. Some of the speed AI gives you leaks out right here. What's left is still faster than writing from scratch.
It has blind spots, too. It won't save a piece that's accurate and boring, or one built on the wrong premise. Those get caught earlier, by judgment.
Anyone can copy a prompt. Nobody can copy the part where I open the pricing page, read it slowly, and catch the thing the draft got wrong.