Share with your CTO
Boris Cherny, the creator and head of Claude Code at Anthropic, is pushing back hard on the prompt engineering industry that has grown up around AI tools. In a Y Combinator interview, Cherny argued that the real constraint on Claude Code’s performance is not prompt construction but verification: giving the model a way to check its own work. He demonstrated this by rewriting the Claude desktop app from Electron to Swift using a simple, direct prompt, no special syntax, no elaborate constraints, and watching Claude self-document its progress in a Slack channel it created on its own.
What this means for your business
Your engineering organization has almost certainly accumulated prompt libraries, internal style guides, and possibly dedicated “prompt engineer” headcount. Cherny’s argument is that most of this investment is fighting the model rather than using it. The teams getting the most out of Claude Code are not the ones with the most elaborate instructions. They are the ones who give the model a hard task, a verification mechanism, and room to work. That is a meaningful reallocation of where engineering effort should go.
The verification point is the one worth taking seriously. Cherny identifies it as “the single most important thing that people do not get right.” What he means is that agents, including coding agents, can run in circles without a feedback loop that lets them confirm progress. Pixel-by-pixel screenshot comparison in the Swift rewrite example is not a clever trick. It is a structured test harness. Organizations that have invested in automated testing infrastructure are already positioned to extend that infrastructure to AI agents. Organizations that have not are running agents blind.
The recurring failure mode Cherny describes, experienced engineers over-specifying because that is how deterministic systems required, is a real cultural problem for engineering leadership. Unlearning it is not a training module. It is a management problem. The signal worth watching is whether your senior engineers are treating Claude Code as a junior colleague to be directed or as a system to be controlled to the instruction level. Those two postures produce measurably different output quality, and the gap will widen as model capability compounds.
Concept deep-dive: Model elicitation
Model elicitation is the practice of surfacing capabilities that exist in an AI model but are not automatically expressed without the right conditions. It exists because large language models are trained on vast, heterogeneous data and develop latent abilities that no single prompt trivially accesses. Think of it like a skilled contractor who can frame a house, do finish carpentry, and wire outlets, but will only do what you explicitly ask unless you invite broader judgment. For enterprise deployments, elicitation matters because the ROI ceiling on a model is set not by the model itself but by how well the organization has learned to surface what it can actually do.
Based on reporting from Head Of Anthropic’s Claude Code Says Prompt Engineering Not That Important, originally published 2026-07-30 08:30:00.
