Boris Cherny’s Advice to AI Users: Focus on Outcomes, Not Steps

Boris Cherny, the creator of Anthropic’s Claude Code, has a message for anyone still writing exhaustive step-by-step instructions for AI models: stop. Speaking at a Y Combinator event, Cherny said one of the most common mistakes he sees is users who underestimate what modern models can handle, micromanaging every detail. “They’re like, ‘I want you to do this, but I want you to do it in this way, this way, this way. You must do one, then two, then three, then four,’” he described.

For today’s systems, that approach is increasingly counterproductive. Cherny instead recommended describing the desired outcome clearly, setting guardrails, and defining what a successful result looks like — and then stepping back. “Then just go let the model cook and come back in a little bit,” he said. “I think it’ll surprise you.” He noted that the same advice would not have worked six months ago, a sign of how rapidly the underlying models are advancing.

The idea echoes what Google Brain co‑founder Andrew Ng called “lazy prompting” last year: providing minimal instructions and adding detail only when the model proves it is needed. Cherny added that even Anthropic’s own employees are still unearthing capabilities in the company’s models, estimating there are likely dozens — perhaps hundreds — of useful abilities that have yet to be discovered.

How ‘Lazy Prompting’ Reflects the Rapid Maturity of AI Coding Tools

Why ‘Lazy Prompting’ Marks a New Phase in AI‑Assisted Development

For years, the dominant wisdom held that getting good output from large language models required precise, chain‑of‑thought prompting — spelling out every sub‑task. That guidance was born from models that often needed hand‑holding to stay on track. Cherny’s comments, grounded in the performance of Claude and its peers, signal that the ground has shifted. When a model can reliably infer the best way to achieve an outcome, the developer’s job morphs from scriptwriter to editor: define the goal and the boundaries, then evaluate what the model delivers.

This is not just a usability tip; it reflects the acceleration of model capabilities. Cherny’s point that his advice would have been wrong half a year ago frames the change as a genuine capability leap, not a stylistic preference. As models become more agentic — capable of reasoning about solutions rather than merely executing a prescribed sequence — the bottlenecks move from prompt construction to task definition and verification.

The Double-Edged Sword of Greater Autonomy

Handing more agency to an AI coding assistant can yield faster, sometimes more creative results, but it also raises the stakes for oversight. Cherny’s emphasis on guardrails and success criteria is critical: giving the model freedom on the “how” only works when the boundaries are explicit. A developer might discover an elegant solution they hadn’t considered, or they might get a path that technically meets the goal but violates an unstated architectural constraint.

This dynamic mirrors the classic management challenge the analogy invites: trusting a capable employee to choose the method while being clear about non‑negotiables. In software, the “employee” is improving every few months, meaning teams must continuously recalibrate how much autonomy to grant. Anthropic’s own engineers, by Cherny’s account, are still finding fresh skills in their models — a hint that today’s best practice may once again be outdated sooner than expected.

What Development Teams Should Change When Working with AI Coding Assistants

  • Replace step‑by‑step prompts with outcome descriptions. Instead of “first parse the JSON, then extract the user ID, then query the database,” write “given this input, return the user’s last three orders in a table, handling any missing data gracefully.” This mirrors Cherny’s advice to describe the task and define success, leaving the implementation path open.
  • Set explicit guardrails without prescribing the method. Clearly state constraints — such as “do not modify the existing database schema” or “only use the approved library set” — but resist the urge to list the steps the model must follow. This balances freedom with safety.
  • Re‑evaluate prompting habits from just a few months ago. Cherny’s observation that the approach wouldn’t have worked six months ago means teams should regularly test whether older, highly detailed templates are still necessary. Run a sample task with a minimal prompt and compare the result to the scripted version.
  • Budget time for review when the model surprises you. An unexpected solution can be a valuable discovery, not a failure. If the outcome meets the guardrails and success criteria, investigate the approach — it may surface a capability you weren’t aware of, aligning with Cherny’s point that many model abilities remain undiscovered.