The most common mistake Boris Cherny sees from Claude users isn’t a bad prompt. It’s too much prompt. Speaking at a Y Combinator event on Saturday, the creator of Anthropic’s Claude Code said people routinely hand the model a numbered march order—do this, then this, then this—and then wonder why the result feels boxed in. His diagnosis is blunt: that style of instruction is a leftover habit, and today’s models don’t need it.What he recommends instead reads like something out of a management handbook, and that’s roughly the point. Describe the task. Set the guardrails. Say clearly what a finished job looks like. Then get out of the way. “Then just go let the model cook and come back in a little bit,” Cherny said, adding that the output tends to surprise people who try it.
Why Claude Code creator says over-specifying every step is the most common Claude prompting mistake right now
Cherny’s argument is that users are underestimating the size of the task a modern model can take on. When someone dictates the method as well as the outcome, they’re spending their prompt on the part the model is already good at figuring out, and spending none of it on the part that actually matters—what success looks like. Be specific about the destination, he says, and loose about the route.
AI chatbot prompting tips age fast, and Boris Cherny says what worked six months ago may be working against you now
Cherny was candid that this approach has a shelf life attached to it in both directions. His own advice, he said, probably wouldn’t have held up six months ago. The models weren’t there yet. That makes prompting less a fixed skill than a moving one, where the techniques people carefully learned in 2025 quietly become friction in 2026.He isn’t alone in the read. Google Brain cofounder Andrew Ng made a similar case last year under the label “lazy prompting”—the idea being that detail gets added to a prompt only at the point where it’s actually required, not pre-emptively.
Beyond prompting: Cherny’s take on how companies should actually measure AI returns
The advice lands alongside a separate framework Cherny laid out in a series of X posts earlier this month, aimed at enterprises trying to justify their AI bills. Usage dashboards are worth watching, he wrote, but they track activity rather than return. The sharper question is whether the company would have put engineering hours into that task anyway—and what those hours would have cost. The larger gain, he added, arrives once fixing and maintaining runs quietly in the background and teams move on to work that wasn’t feasible before.That’s the mood across the industry now. The token-burn arms race of early 2026 has cooled, with Jamie Dimon and Sam Altman both flagging cost-versus-value as the question of the moment. Cherny reckons dozens, possibly hundreds, of useful model capabilities are still undiscovered—but you only stumble into them if you stop dictating the route.
