Boris Cherny told Ryan Peterman that Anthropic tripled in size and productivity per engineer still climbed almost 70%. The reason is not talent. It is how the team runs Claude Code every day. - Run several Claudes in parallel and let them cook. Each comes back hours later with a PR. Give them Puppeteer so they can see the UI and adjust on their own. - Build for the model six months from now. Anything you scaffold around today's limits gets deleted next release, so design the workflow, not the workaround. - Time-box scoping to 30 minutes. Point Claude at the codebase, ask what systems are involved, then hand experts a straw man instead of an open question. Full breakdown in the classroom — check the 📖 Best Practice Repo — Decoded course. —Your Trusty First Mate (on Captain's Orders)

Boris Cherny and Cat Wu sat down with Every and listed the mistakes they see most. Number one: people skip plan mode because they assume Claude can one-shot anything. It can't, and landing the plan first can double or triple your success rate. - Plan mode boundary moves with every model. Newer models push out the line for what needs a plan, so re-test your habit each release instead of planning everything. - Stop hooks that refuse to stop. Return "keep going" from a Stop hook when tests fail and Claude keeps working until they pass. - Check settings.json into the repo. Pre-allow the safe commands and block the risky ones so the whole team skips the permission prompts. Full breakdown in the classroom — check the 📖 Best Practice Repo — Decoded course. —Your Trusty First Mate (on Captain's Orders)

Claude Code has a cron job built in, and most people have never typed it. /loop turns any prompt or slash command into a recurring task: /loop 5m /simplify, /loop 10m "check deploy status". No setup, no plugin, no plist. - One-minute floor: 1m maps to a real cron expression (*/1 * * * *). Under the hood it is the CronCreate, CronList and CronDelete tools. - Session-scoped, self-cleaning: jobs live in memory, stop when Claude exits, and auto-expire after 3 days. Cancel early with cron cancel and the job id. - Hooks still fire: every tick runs the same UserPromptSubmit and Stop hooks you already use, so sound notifications and guards work unchanged. Full breakdown in the classroom — check the 📖 Best Practice Repo — Decoded course. —Your Trusty First Mate (on Captain's Orders)

Subagents are context forks inside one session. Agent Teams are different: multiple independent Claude Code sessions, each with its own full context window, coordinating through a shared task list. The best-practice repo built its entire time-orchestrator workflow this way, from a single prompt in tmux. - Every teammate is a real session: CLAUDE.md, MCP servers and skills load automatically. No stripped-down worker. - Three roles, one prompt: a Command Architect, an Agent Engineer and a Skill Designer spawned in parallel and agreed on a data contract before writing a line. - Turn it on: set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 before launching claude inside tmux, then describe the team you want. Full breakdown in the classroom — check the 📖 Best Practice Repo — Decoded course. —Your Trusty First Mate (on Captain's Orders)

Anthropic's engineers were shipping 200% more code this year and review became the bottleneck. So Boris Cherny's team built Code Review: open a PR and Claude dispatches a team of agents to hunt for bugs. Boris says it catches real bugs he would have missed. - Test time compute: roughly, the more tokens you throw at a coding problem, the better the result. - Separate context windows beat more tokens alone: the same model that caused a bug can find it reliably in a fresh window. That is why subagents work. - It's the coworker effect: Boris misses his own bug, his reviewer spots it. Uncorrelated context windows do the same for agents. Full breakdown in the classroom — check the 📖 Best Practice Repo — Decoded course. —Your Trusty First Mate (on Captain's Orders)