Or rather, you're prompting it at all when you could be orchestrating it. Boris Cherny built Claude Code. In fifteen minutes he shows you the actual stack he runs daily, plan mode, CLAUDE.md, subagents, hooks, worktrees. Watch it once. Send it to your team.
Ike Singh Kehal posted a clip this week from Boris Cherny, the engineer who built Claude Code at Anthropic. Made it mandatory viewing for his team. I watched it twice. He's right.
Every developer I know has Claude Code installed. Almost none of them are using it the way Boris does. That's not a complaint about the developers. It's a comment on the product. The defaults assume you'll figure out the rest. Most people don't.
What follows is fifteen minutes of him showing you the actual stack. Plan mode. CLAUDE.md. Subagents. Hooks. Worktrees. Watch it once.
The stack
I'll do you the favour of writing it out, so you can refer back without scrubbing.
Plan mode is not optional
Stop prompting from scratch. Have Claude write the plan first, edit it, then execute. The plan is the brief. Skipping it is the most common mistake. Most people are writing one-shot prompts when they could be writing one-shot briefs. Different sport.
A real CLAUDE.md, not an empty one
Every session starts cold without it. Put your conventions, your gotchas, your repo geography, the lib you don't use, the lint config that bites people. That's the file Boris keeps pointing back to. Not a config trick. A brief for your repo.
Subagents do the boring parts
Spin them up for parallel work. Tests in one. Linting in another. Investigation in a third. You stay on the main thread, thinking about the actual thing. The subagents come back with answers. That's not autocomplete. That's a team.
Hooks make it yours
Tie Claude into your repo lifecycle. Auto-run tests on save. Auto-format on commit. Cherny runs a stack of them. By the time you've configured three of these, the tool stops feeling like a chat window and starts feeling like part of your environment.
Worktrees so three Claudes ship three branches
Spin up parallel git worktrees, run a Claude in each, ship three branches in the time you'd usually do one. This is the moment where the framing of "using AI" stops fitting. You're not using one engineer. You're directing several.
Why this matters
I keep saying the brief is the new interface. The shift in this video is the same shift Anthropic shipped with Claude Design and Figma shipped with their canvas agent. The thing you write is the thing that gets executed. The tool becomes whatever shape the brief requires. The job is the framing, the edit, and the taste.
The clip below isn't about Claude Code. The clip is about what happens when the person who built the tool sits down and shows you the tool. Most software stops there. Boris doesn't. He shows you the bench he runs around the tool. That's the bit most software demos skip.
The gap between "I use Claude Code" and "Claude Code is my best engineer" is one video.
Watch it once. Send it to whoever on your team has been using Claude Code as a slightly smarter autocomplete. Then watch their next sprint.