The best Claude users are good managers

Anthropic studied about 400,000 Claude Code sessions and its own staff. The people who got the best results weren't the most technical. They briefed Claude clearly, fixed their prompts instead of patching outputs, made results easy to review, and kept their setup lean.

Notes on Ben's video · watch on YouTube · click any timestamp to jump there

Beginners vs experts

"Expert" here means the clearest communicator, not the most senior person. The study rated users by how precisely they described the task, what they asked Claude to check, and who corrected whom.

Got a working, verified result
Beginners 15% · experts over 2× 0:54
Gave up when something went wrong
Beginners 19% · others 5–7% 5:24
Still reached the goal after a failure
Experts 4× as often 5:46
Decisions the best users keep for themselves
~70% of planning · ~20% of execution 2:52
Length of Claude's replies
Beginners ~600 · experts ~3,200 words 8:53
BeginnersExpertsFigures as reported in the video; I haven't checked them against Anthropic's papers.

Brief Claude the way a manager briefs a new hire

Cover these four parts in every prompt. Click one to see it in the example. 2:36

The six lessons

Pick a lesson to open it.

Lines and examples worth keeping

What I'll change in my Claude Code work

Tick these off as they become habits. Ticks are saved in this browser.

    Check what you remember

    Known: 0 of 5
    Transcript fixes and open questions
    • The auto-captions garbled the names. "Boris Journey" is Boris Cherny, who created Claude Code. "Amanda Ascol" is Amanda Askell, and "Andre Karpathy" is Andrej Karpathy.
    • The transcript's "/re" is most likely /rewind. Pressing Esc twice does the same thing.
    • The intro's claim that experts "get more than 50% worked on" doesn't parse, so I left it out.
    • The output-style command reads "/output style concise include code". Check it in the video at 9:13.
    • I skipped the sponsor segment (11:31–12:04) and the free-resource plugs.