Essay · How I work
The first draft always reads finished
What bullshots taught me about AI drafts, and the three passes I run on every one.
I watched E3 2005 as a player, years before I had any job in games. Sony ran a Killzone 2 trailer that looked like nothing a console had ever run, and its people told the press it was real. (Source ) It was a target render, the studio's picture of what it hoped the PlayStation 3 could do. (Source )
That September, Penny Arcade handed everyone a word for marketing that looks finished and isn't: the bullshot. (Source ) Players ran with it. I started writing about games for Joystiq in 2007 and went to E3 myself in 2010, for GamePro. I've been squinting at things that look finished ever since.
An AI first draft is a bullshot you ordered yourself. It has the structure and the confidence of finished work, and nothing on the surface tells you which parts run on the real engine.
Heroic fantasy is a hard place to be modest
Late in 2010 I crossed to the studio side and wrote launch email for Lord of the Rings Online and Dungeons & Dragons Online at WB Games Boston. You're telling someone a new feature will make them a rider of Rohan, or the wizard the party can't live without. A lot of the time the game could deliver that. A big part of my job was catching the sentences where it couldn't, after enough hands had turned a plain product description into a promise.
Trust spends like any other resource, and no patch refunds it.
Ask Sony. In May 2024 it told Helldivers 2's PC players they'd have to link a PlayStation Network account. Steam took more than 200,000 negative reviews in three days, (Source ) and Sony dropped the requirement. (Source )
The writing got good, and that's the trouble
To be fair to the models, they write well now. In a New York Times blind quiz in March 2026, 54% of more than 86,000 readers preferred the AI passages. (Source )
Clean prose stopped being evidence of anything.
Accuracy has improved too, though any number I gave you would be a choice. Summarizing a document it's been handed, the best model on one leaderboard makes things up under 2% of the time. (Source ) On a hard knowledge test in Stanford's 2026 AI Index, the best of 26 models still had a 22% hallucination rate. (Source ) They measure different things, and both will be out of date soon.
The part that doesn't age is the gap between what a model can read and what only your company knows. Say you ask for a case study and don't supply the results. You might get this: “Customers cut onboarding from nine days to two.” It reads like proof, and nobody measured it. (I invented that one. A draft can do the same and not tell you.)
Damien Charlotin's public database lists more than 2,200 court decisions where someone relied on AI-invented material. (Source ) More use explains part of the climb, so I read it as a record of people who never opened the source.
Better models still can't check what they were never told
Newer models catch more of their own mistakes, and search helps. What they can't check is what never reached them, like the thing sales promised last quarter. Somebody has to walk in with the source document, the pricing page, and a real sense of who's reading.
Most of the work happens before the prompt
I pull the sources, ground the model in them, and go looking for the product's limits, usually features locked behind a higher tier or places where it's weak. Then I tell the model to flag what's missing instead of filling the gap.
Three passes, claims first
- Pass 1 · Check the evidence
I open the sources and check every claim. Even a real number often 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. Rider of Rohan is fine if the horse shows up.
- 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.
What the same hours buy now
People assume AI means less time on the piece. Mine take about as long as they used to. What changes is what fits inside those hours: more research, more rounds of refinement, marketing and design and coding agents all working the same problem. It's closer to having a full team in the room. Every role gets covered, and somebody still has to run the meeting.
That somebody is a person, often more than one, before anything ships. Run it carelessly and all you've built is a bigger bullshot.
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.