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. The job's core mechanism was catching copy promising something beyond the game's reality. A sentence would drift through one too many people in different parts of marketing and production, and what started as a clear product description would become a grand abstracted promise we couldn't quite deliver.
Where the drafts break
A bad forgery gets stopped at the counter. A good one? It can circulate for years. The cost hits the center of whoever holds that forgery when an authority finally verifies its authenticity.
AI drafts fail by sounding right. It just takes a while after you hit publish.
A model will write “cuts research time per prospect from 20 to 2 minutes” because that sentence has the shape of a claim a buyer responds to, and shape is what AI is good at. The problem is that shape tends to be the same one over and over and over. It 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% carried at least one significant issue, and 20% had major accuracy problems including invented details. And that's the easy version of the job. The model had the article sitting right there.
The draft also arrives with evened out paragraphs, like a wax cake, with a confident topic sentence at the head of every section. Those top-level sentences used to mean somebody spent time, effort, and mental focus. Now they cost little to nothing.
Why the model can't check itself
Ask the model to check its own claims. It'll tell you they're fine.
Somebody actually tested this.
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 of knowing whether a claim is true. Ask it to grade its own work and you get a second confident answer, not a check.
Somebody has to walk in with what the model doesn't have: the source document, the pricing page, the product philosophy, the emotional center of audience knowledge.
Before I prompt
More than half the work happens before a draft. I ask myself: Who is this for? What does it have to do? I'll gather sources. I'll search for the product's limits. This typically means features hiding behind a higher pricing tier and features where the product is very weak. Then I tell the model to flag what's missing instead of filling the gap.
That last instruction helps, but it's not enough. A model told to flag gaps will still write a confident sentence followed by another sentence telling you why that previous sentence is so important.
What I check, and in what order
The review pass
- Pass 1 · Check the evidence
Open the sources. Check whether the claims hold up. A number like 20 to 2 minutes usually traces back to one customer on one call. One customer's result isn't what everyone should expect, so I narrow the claim to what the evidence actually supports, or I cut it.
- Pass 2 · Check the promise
A promise has to describe what the product actually does, even if that promise is adorned with the wings of high emotion.
- Pass 3 · Check the claim, then the prose
Do the facts support the claim? Have I left out anything the reader needs? Then rhythm and specificity, including whether a competitor could use the same sentence unchanged. All this helps rank what words get higher priority, what becomes a header, and what words and even whole paragraphs get cut.
Claims come first: removing one will often change everything around it.
Overclaiming is the obvious failure; hedging is the quiet one. “Can help reduce X” can't be proven wrong, which means it isn't really a claim, and readers price it that way. If the evidence supports “does,” then write “does.” The sentence gets shorter, its promise gets stronger, and now it's something a customer could hold you to. If you deliver? That's a win on brand prestige, repeat customers, and word of mouth.
Then I read the finished copy once more for accuracy and overall flow, headline included. Headlines are where claims get compressed hardest and checked least. That flow portion is more of a feeling, less of an intellectual decision. Great writing is like great music. There are trends, there are fundamentals, and there are rules to know so we know when to bend or break them for maximum results.
What it costs, and what it misses
The whole pass takes time and energy, both mental and emotional. On a long piece it can take longer than the draft did, because checking a source is slower than reading a sentence. Weighing a feeling is slower than drafting a hopeful promise. The speed AI gives you is real, and some of it leaks through the cracks inherent to the AI limitations. What's left is still faster than writing without AI involved, but intuition cannot be modeled by AI. Intuition is where the best writers thrive.
The pass doesn't catch a piece that's accurate and boring, a claim built on the wrong premise, or a subject nobody should have written up in the first place. Those are earlier problems, avoided through intuition and context clues wrapped up in the white-water rapids of human emotion.
Why it holds
Prompts, even agentic workflows, are easy to write and easy to copy. Anyone can re-use a prompt, a workflow. Nobody can repurpose that intangible intuition we all have inside, where I open the pricing page, read it, and silently reflect until that brilliant eureka moment.