Boris Cherny says there is no secret to prompting
Talk to Claude like a coworker, say what you want, how much effort to spend, and how it should check its own work.
Boris Cherny posted his actual prompts after people reacted to how plain they were, and said the reaction itself surprised him. His advice is to talk to Claude the way you would a coworker. There is no secret to prompting, and for most tasks there is no need to be heavily scaffolded or prescriptive.
There's no need to be overly scaffolded or prescriptive for most tasks -- give Claude a goal, and it will figure it out.
What he says does matter is three things. One, what you want it to do. Two, how much effort you want it to spend. Three, how it should verify that it did the right thing. That is the whole list.
The prompt rules got shorter as the models got better
Cherny frames this as a change over time rather than a universal truth. Back in the Sonnet 3.5 days, he says, your prompt mattered a lot. The implication is that a lot of the technique people built up then was compensating for something the models no longer need help with, and that carrying it forward is just overhead.
That lines up with the argument dax made earlier about elaborate custom agent setups fixing problems that no longer exist. Same shape, different vantage point, and in Cherny's case it comes from someone whose prompts people were curious enough about to ask for in the first place.
The third item is the one that does real work and the one most people skip. Telling the model what it wants is easy. Telling it how to confirm it got there, which command to run, which output to look at, what a pass looks like, is the part that turns a plausible diff into a checked one. The effort dial is the other underrated one, because the same sentence can mean a quick patch or an afternoon of searching depending on what you say about it.
This is advice, not an experiment. Cherny did not post a benchmark or a before and after, he posted screenshots of what he actually types and an explanation of why it is so short. Treat it as what one practitioner does, and notice that what he does is less than what most prompt guides recommend.