What Anthropic's Own Teams Do Differently in Claude Code

Anthropic published a write-up on how its own teams use Claude Code, and the detail most people miss is who got the biggest wins. Not the senior engineers. The legal team prototyped an internal phone-tree tool with zero dev resources. Growth marketing built an agentic pipeline that turns a CSV into hundreds of ad variations in minutes — plus a Figma plugin that cut hours of copy-pasting down to roughly half a second per batch. RL data scientists who don't write TypeScript now ship whole React dashboards to visualize model performance.

Meanwhile, the engineering teams' habits all collapse into one principle. If you only steal one thing from the people who built the tool, steal this: never end a prompt without telling Claude how to prove the work is correct.

1. Give it a check it can run

Claude stops when the output looks finished — and without a pass/fail signal, you are the test suite. Anthropic's best-practices guide is organized around closing that loop: hand Claude a test command, a build step, a lint pass, or a reference screenshot to diff against, and demand the evidence (test output, command results), not a claim of success. Their security engineering team credits this shift — plus moving to test-driven development with Claude — for resolving incidents that used to eat 10–15 minutes of manual log-scanning about 3x faster. For unattended runs, a Stop hook can block the session from ending until the check passes (Claude only overrides it after 8 consecutive blocks).

2. Plan only when the diff needs a paragraph

Plan mode makes Claude explore the codebase and produce an implementation plan before touching anything, and Ctrl+G opens that plan in your editor so you can tighten it by hand. But the docs are explicit about when to skip it: if you can describe the change in a single sentence — a rename, a log line — just ask directly. Planning earns its overhead when the work touches several files at once or you haven't settled on an approach yet.

3. Treat CLAUDE.md like code, not a diary

The pruning test from the docs is brutal and useful: for every line, ask would removing this cause Claude to make mistakes? If not, cut it. An overgrown CLAUDE.md actively backfires — important rules drown in noise and Claude starts ignoring them. Review it when things go wrong, prune it regularly, check it into git so the whole team compounds on it.

4. Two failed corrections, then /clear

Context is the real constraint: performance degrades as the window fills. Anthropic's rule is refreshingly concrete — if you've corrected Claude twice on the same issue in one session, stop correcting. The context is now polluted with failed attempts. Run /clear and restart with a sharper prompt that bakes in what you learned. Subagents follow the same logic: research runs in a separate context and reports back a summary, so your main thread stays clean for implementation.

5. Point it at problems outside your stack

This is the underrated pattern. Anthropic's data infrastructure team debugged a Kubernetes pod-scheduling failure during an outage by pasting dashboard screenshots — Claude walked them through Google Cloud's console and produced the exact commands to create an IP pool, saving about 20 minutes mid-incident. The inference team writes Rust tests without knowing Rust by describing the behavior they want. The ceiling is no longer your strongest language; it's how precisely you specify and verify.

Why this matters for your career

For a Brazilian dev or EM aiming at international teams, this write-up reads like a job description. Companies hiring globally increasingly assume you already work this way — and the person who sets up the verification loop, the lean CLAUDE.md, and the review workflow for the whole team is doing visible, promotable leadership work, whatever their title says.

Today: take the next task you'd hand Claude Code as "implement X" and rewrite it with a verification clause — the test cases, the command to run, and "iterate until it passes, then show me the output." Compare that session with your usual one.

Sources

• Anthropic — How Anthropic teams use Claude Code
• Claude Docs — Best practices for Claude Code