A book, and three free pieces from it · Buy the book →
How to get real work out of AI — and keep getting it when the models change.
If you've read anything about prompting, you've been told three things: give the model an expert persona, be polite to it, and tell it to think step by step.
All three have been tested.
162 role assignments across 2,410 questions and four model families. No consistent improvement; the average effect was slightly negative. Low-knowledge personas often reduced accuracy outright.
The finding sits at an angle to the advice: there's no bonus for courtesy, but there is a penalty for rudeness — and even that fades on modern models. Keep the habit for your own sake if you like, just not because it buys better work.
Excellent, well-evidenced advice in 2022. OpenAI's own guidance now says it's unnecessary for reasoning models, and that such techniques "may not enhance performance (and can sometimes hinder it)." Wharton measured one reasoning model where it was slightly negative.
Nothing went wrong here. The last one was genuinely good advice — a real finding, a real effect. It didn't fail. It got absorbed: what was once an external trick became a behavior trained into the model itself.
That's the pattern worth having. The trick expired; the principle didn't. Anyone who memorized the words has to relearn. Anyone who understood why it worked changed nothing at all.
Now consider what that implies about a book selling you three hundred prompts.
In a randomized trial run in early 2025, sixteen experienced developers using AI took 19% longer on real tasks in code they knew well. Hold that number loosely, though not for the reason you'd guess. METR re-ran the experiment with 57 developers and reported a problem with the new run: people who won't work without AI increasingly decline to take part, which biases the newer measurement against showing a benefit. They think AI probably helps more in 2026 than it did in 2025, and that their own data is too weak to say by how much. Nothing was retracted. The book goes through what they actually said, because the version everyone repeats is a retraction that didn't happen.
What they believed is the part that survives the revision. Beforehand they expected to be about 24% faster. Afterward, having just done the work, they still estimated they'd been 20% faster.
Their perception didn't merely miss. It had the wrong sign, and direct experience did not correct it. A correction to how much slower they were doesn't touch the fact that they couldn't feel it.
Sixteen people can't carry a general claim, and the book doesn't ask them to — that job belongs to several decades of automation-bias research finding the same under-monitoring in experts and novices alike, in cockpits, control rooms and clinics. This trial is the vivid version, not the evidence.
The hard part isn't getting AI to produce something. It's knowing when what it produced is worth having — and you cannot rely on your own sense of how well it's going.
That's not a prompting problem. No phrasing solves it.
The context pack template from the book's appendix — the single change that improves most people's output today. Plus the three posts as they publish. No other email.
It assumes you already use AI. No tour of what a language model is, no convincing you it's useful, no prompt library that expires next quarter.
It's a map of where AI genuinely helps and where it quietly costs you, six systems you install once, a verification method fast enough that you'll actually keep it, and an argument about what you should never hand over.
Every system is demonstrated end to end, including where the demonstrations failed. You'll watch a model invent a launch date it was never given, promote a passing remark into a tracked commitment, and quietly resolve a question that was deliberately left open.
You'll also watch one chapter's central failure refuse to reproduce under testing — and the chapter report that, rather than quietly keeping the better story.
Written with AI, disclosed up front. The research, drafting, and revision were done by a language model working from direction. Every technique in Part II was demonstrated rather than described. The scenarios those demonstrations run on are constructed, and the chapters using them say so; the model's output is verbatim. Where a demonstration failed, the failure is reported, including the one that could not be reproduced at all. Chapters that make factual claims list their sources at the end, most of the practice chapters carry none, and where the evidence is weak the book says so rather than rounding up. A book about getting real work out of AI seemed like the wrong place to be coy about how it was made.
Kindle eBook, $9.99, or paperback, $19.99. Both on Amazon, both KDP Select.
Prefer paper? Buy the paperback ($19.99).
Everything the book tells you to build, already built: the context pack with four filled role examples, the six systems from Part II as standalone playbooks, the rules page, the tool-evaluation worksheet, and the quarterly review. Plain text, no vendor lock-in.
What's in the System Pack → — finished, priced, and not on sale until the checkout is wired. The list hears first, at a lower price than it launches at.
The six systems, installed in your team in one sitting. Everyone leaves with their own context pack written, one system adapted to work they actually do, and a team rules page saying what gets verified and what never gets handed over.
Formats and prices → — from $6,000 for three hours remote.