Claude Code's creator says every prompt needs a proof step
Boris Cherny's three-part prompt advice ends with verification. In finance, that third part is the control.
What happened
Boris Cherny, who created Claude Code at Anthropic, posted prompting advice on X on October 6, 2026, Business Insider reported. He said prompts no longer need heavy scaffolding. “Talk to Claude the way you would a coworker,” he wrote. He named three things to include: what you want done, how much effort to spend, and how the model proves it did the right thing. He also said he trusts Anthropic’s Opus 5.5 on matters of taste. “I trust it to cook,” he wrote.
What it means
The first two parts describe the task. The third describes the evidence. A prompt that says how to prove the result is a control written in plain language. Finance already works this way. A reconciliation is done when it ties to the bank statement, not when someone says so.
Why it matters to you
Finance teams that write prompts for close, cash application, or collections can borrow the format as is. State the outcome. State the effort. State the proof.
One example is “tie the cash applied to the bank deposit total and list any difference.” An agent that must show its proof is easier to review and easier to reverse. Verification also makes reversal easier. When the proof fails, the reviewer knows exactly which step to undo.
What to watch
Watch whether Anthropic builds a proof step into Claude Code itself, so the model shows its evidence without being asked. Watch whether other AI labs give the same three-part advice.