There’s a moment most people have with AI where they decide, quietly, that it’s overrated.
They asked it to write something — an email, a summary, a plan — and what came back was fine. Generic, but fine. Competent in the way a stock photo is competent. Not wrong, just nobody’s.
So they concluded the tool was the problem.
It wasn’t. The problem was that they asked a stranger to do their job.
Here’s a task almost everyone has had: tell your team a launch is slipping two weeks.
Cold, with no other information:
Write a short note to my team announcing that we’re pushing the launch date two weeks.
What came back:
Subject: Important Update Regarding Our Launch Timeline
Hi team,
I hope this message finds you well. I wanted to reach out with an update on our upcoming launch.
After careful consideration and thorough evaluation of our current progress, we have made the decision to move our launch date back by two weeks. While this adjustment may come as unexpected news, we believe it represents an exciting opportunity to ensure we deliver the highest quality experience to our users.
I want to thank each of you for the incredible dedication and hard work you have poured into this project…
136 words. It opens with “I hope this message finds you well,” which nobody has meant since about 1997. It describes a delay as “an exciting opportunity.” It thanks the team twice.
And — the important part — it never says why the launch is slipping, because it doesn’t know. So it gestures at “careful consideration” instead.
That’s the whole problem in miniature. The model had a hole where the reason goes. It filled it with the shape of a reason.
Now the same request with context. I changed nothing about the sentence. I added three things first:
Same model. Same request. Minutes apart.
Subject: Launch moving to the 24th
The payments vendor’s sandbox slipped and we can’t test checkout without it, so launch moves two weeks to the 24th.
Nothing’s wrong on our side — this isn’t a quality call or a scope change. I’d rather tell you it’s a vendor dependency than dress it up.
Dates for the revised plan go out tomorrow. Anything already in flight keeps its current deadline.
62 words. Reason in the first sentence. No apology for something that doesn’t require one. It anticipates what the audience was predicted to assume and kills it. It ends on what happens next.
Nothing about the second request was cleverer. It was better-informed.
Look at that subject line again: Launch moving to the 24th.
I never supplied a date. The model inferred one, stated it twice, confidently, in a message about a commitment.
Sent unread, the team now has a date that came from nowhere. It’s fluent, specific, plausible, and invented.
Both things are true at once: the output was dramatically better and it committed me to something I hadn’t decided. That’s the actual texture of this work, and anything showing you only the clean version is selling you something.
The clearest evidence comes from research on how these models learn from examples.
Give a model a few demonstrations of a task and it performs better — long established. The interesting part is why. Researchers replaced the correct answers in those demonstrations with random ones and found performance barely moved.
What mattered wasn’t whether the examples were right. It was that they conveyed the shape of the task: the format, the kind of input, the space of possible answers.
Demonstrations work by supplying context and shape, not by instructing.
Which is exactly why pasting two real emails beats writing “professional but warm.” The adjective describes a target. The examples supply the thing itself.
Make one document — context.md, a pinned note, whatever
you’ll find again.
1. Who you are, in work terms. Role, what you’re responsible for, who your output goes to. Two or three sentences. Not a bio.
2. Your voice, shown not described. Paste two or three things you actually wrote — a real email, a Slack message, a paragraph you were happy with. Include at least one short one; your chat register is more clipped than your email register, and a model given only formal samples will make everything formal.
Don’t write adjectives here. “Professional but warm” tells the model nothing it can act on.
3. The people and things you reference constantly. Your team, your stakeholders and what they care about, the projects you refer to by shorthand. This is what stops you re-explaining who “Dana” is every single time.
4. Your standing constraints. These four are worth stealing directly:
- Never commit me to a date, number, or scope I haven’t stated.
- Flag anything you’re inferring rather than drawing from what I gave you.
- Don’t add points I didn’t include — list them separately at the end.
- Don’t resolve questions I left open. If I hedged, keep the hedge.
That first line is the one that would have caught the invented date above.
Then use it for a week. Start every substantive session by supplying the pack before the request. Watch the first reply — before you’ve done any clever prompting at all.
I should be honest that no single study directly pits “richer context” against “cleverer phrasing” and measures which contributes more. This is a synthesis across several literatures, not a finding I can point to.
I’d rather say that than let a confident sentence do work the evidence doesn’t support — particularly in a piece arguing you should verify what AI tells you.
Adapted from Chapter 3 of NO MAGIC WORDS — how to get real work out of AI, and keep getting it when the models change.
Sources: Min et al. 2022 (EMNLP, arXiv:2202.12837 — the random-label finding) · Brown et al. 2020 (NeurIPS) · Lewis et al. 2020 (NeurIPS) · Horvitz et al. 2024 (EMNLP Findings) · Sclar et al. 2024 (ICLR)
The fifteen-minute setup from the book's appendix — the single change that improves most people's output today. Plus the rest of this series, and a note when the book is out. No other email.