How Claude Code's Boris Cherny works with Claude, in his own posts
The head of Claude Code says prompting is now about the goal, the effort and the check, and his public posts show what that looks like day to day.
If you use Claude at work, the clearest advice Boris Cherny has given lately fits in three lines: tell Claude what you want, how much effort to spend, and how it should check that it did the right thing. He's the creator and head of Claude Code at Anthropic, and his public posts from the last six weeks show him working that way every day. This profile is built only from what he has said in public, with a link beside every quote.
Who he is#
The Pragmatic Engineer describes Cherny as "the creator and Head of Claude Code at Anthropic," who "previously spent five years at Meta as a Principal Engineer and is the author of the book Programming TypeScript" (The Pragmatic Engineer, Mar 4). Lenny's Podcast lists, among its episode topics, why he briefly left Anthropic for Cursor and came back after two weeks (Lenny's Podcast, Feb 19). In an Oct 13 interview with Nikkei xTECH, he described Claude Code's growth as hyper-hypergrowth with every resource stretched thin, and said he thinks the team can get past those limits (our paraphrase of the Japanese article, Nikkei xTECH).
What changed in how he prompts#
The before and after is his own. On Oct 6 he wrote: "Back in the Sonnet 3.5 days, your prompt mattered a lot." Now, he wrote, it matters more to tell the model "What you want it to do," "How much effort you want it to spend" and "How it should verify that it did the right thing." His shorter version: "Talk to Claude the way you would a coworker." He added: "There's no secret to prompting." (X, Oct 6)
Why it matters for you#
If your prompts are long lists of rules written for older models, Cherny's advice suggests trimming them. His three-part brief is short enough to use on any task, and the third part, the check, is the one people skip. It's the difference between "write the fix" and "write the fix, reproduce the bug, and show it passing."
What his workday looks like#
- Parallel sessions with a plan first. The Pragmatic Engineer's summary says he works across five terminal tabs, starts Claude in plan mode, refines the plan, then lets it build, quoting him: "once there is a good plan, it will one-shot the implementation almost every time." (The Pragmatic Engineer)
- One running conversation instead of many sessions. On Projects in Claude Code, he wrote: "I stopped managing sessions." (X, Sep 17)
- Claude in Slack doing real work. He wrote: "Tag writes >50% of my PRs every day." One example prompt he shared: "@Claude brainstorm ~100 hypotheses for what could explain this weird data, use a workflow to in/validate. spend 10m tokens or so digging very deep. draw a chart with the result." That prompt has all three parts: a goal, a budget and a visible check. (X, Sep 25)
- A reviewer on everything. "I could not imagine being productive without Claude Code Review." (X, Oct 16)
- Formal methods for the hardest code. He used Lean to check the Claude Agent SDK and wrote: "A couple short prompts = 16 PRs fixing various bugs and race conditions." (X, Sep 22) A day later he clarified: "It's not that the whole codebase is formally verified (yet!..), more that the hairiest parts of the code are modeled, checked for counter-examples, and fixed." (X, Sep 23)
How to use it#
- Write the goal in one or two plain sentences, as if briefing a coworker.
- Give an effort budget, in minutes or tokens.
- Name the check: tests, a repro, or the numbers behind a chart.
- Put a reviewer on every change.
This template is ours, built from his three points, and we haven't tested it:
Goal: <what you want done, in one or two sentences>.
Effort: spend up to <time or token budget>; if it looks bigger than that, stop and tell me why.
Check: before you say you're done, <run the tests / reproduce the bug and show it passing / show me the numbers behind the chart>, and list anything you could not verify.
Good to know#
- His setup isn't everyone's. Projects in Claude Code launched "In beta for select users." (@ClaudeDevs, Sep 17), and Code Review is generally available "for Claude Team and Enterprise plans" (@ClaudeDevs, Oct 15).
- Big budgets cost real usage. A prompt that tells Claude to spend millions of tokens will use your plan's limits. Set a budget you can afford, not his.
- He works at Anthropic. His posts are first-hand accounts of his own use, not independent tests, and he's promoting products his team ships.
What to watch#
Cherny noted on Oct 22 that Agent Skills "started as an experiment in Claude Code" before Anthropic donated the standard to the Agentic AI Foundation (X, Oct 22). That's one habit that started inside his team's tool and became a shared standard. The signal to watch is the next feature he says he uses every day in his own posts, and whether @ClaudeDevs follows with docs and plan availability.
Last checked: Oct 26, 2026, ~1:31 PM ET. What changed: cut YC Startup School block (paragraph, Good to know bullet, YouTube source, step 4); added Sep 23 X post to sources; confidence stays confirmed.
Steps
- Write the goal in one or two plain sentences, the way you would brief a coworker.
- Say how much effort to spend, as a time or token budget.
- Say how Claude should check its own work: tests to run, a repro to pass, or a chart to draw.
- Put a reviewer on every change, whether that's Code Review on Team or Enterprise or a second session that only reviews.
Goal: <what you want done, in one or two sentences>. Effort: spend up to <time or token budget>; if it looks bigger than that, stop and tell me why. Check: before you say you're done, <run the tests / reproduce the bug and show it passing / show me the numbers behind the chart>, and list anything you could not verify. (ours, untested)