How Anthropic's Own Teams Actually Use Claude Code
Most people use Claude Code like a fancy autocomplete. Anthropic's internal teams use it differently, and the gap in results is significant. Here's what 10 teams across the company actually do, distilled into the things worth stealing.
The mindset shift that matters most
Stop treating it as a one-shot oracle. The API Knowledge team's core principle: approach Claude Code as a collaborator you iterate with, not a machine you query. Start with the bare minimum context and let it guide you through the process. The more you front-load, the worse the results tend to be.
The second shift is learning to classify your tasks. Some things you hand off completely (peripheral features, test generation, refactoring, documentation, visual tweaks). Other things you supervise closely (core business logic, security-sensitive code, anything touching production). The product team calls this "task classification intuition" and considers it the most important skill to develop.
Your CLAUDE.md file does more work than you think
Every high-performing team invested heavily in their CLAUDE.md. It's where you document your stack, your workflow expectations, and mistakes you want Claude to stop making. The RL Engineering team added one line that significantly improved consistency: "run pytest not run, and don't cd unnecessarily, just use the right path." Simple. Worked.
The Product Design team told Claude they were designers with little coding experience who needed detailed explanations and incremental changes. That single instruction in CLAUDE.md transformed the quality of responses for their use case.
The self-improving loop: at the end of each session, ask Claude to summarize what happened and suggest improvements to the CLAUDE.md based on actual usage. Over time, the file gets smarter and so does every session that follows.
The hack the Claude Code team itself uses
The team that actually builds Claude Code uses their own product in a specific way worth noting. For peripheral features they don't fully understand, they enable auto-accept mode (Shift+Tab), give Claude an abstract problem, and let it work completely autonomously. They review the output when it's about 80% done and take over from there.
This is how Vim key bindings got built. Nobody on the team prioritized it, Claude was just asked to build the entire feature, and roughly 70% of the final implementation came from that autonomous run. A few iterations to complete it. That's the real workflow for non-critical features.
The important setup: start from a clean git state and commit checkpoints frequently. If Claude goes sideways, you revert and try again. No drama.
Screenshots are criminally underused
Three separate teams independently landed on the same technique: feed screenshots directly into Claude Code.
Data Infrastructure diagnosed a Kubernetes cluster failure by pasting in dashboard screenshots. Claude walked them through Google Cloud's UI step by step until it spotted the warning about pod IP exhaustion, then gave them the exact commands to fix it. No networking specialist needed.
Product Design pastes mockup images straight into Claude Code (Command+V) to generate working prototypes that engineers can immediately build on. Legal does the same thing, showing Claude what they want an interface to look like rather than trying to describe it in text.
If you're not pasting screenshots in, you're skipping one of the most useful things it can do.
The slot machine technique
Data Science and ML Engineering teams use this for refactoring tasks that are too messy for macros but not significant enough for a proper sprint.
Commit your current state. Let Claude run for 30 minutes. Either accept the result or start fresh.
Don't try to fix a bad trajectory. Starting over genuinely has a higher success rate than trying to correct Claude's mistakes mid-run. The commit before you let it loose is what makes this safe.
Plan in Claude.ai, build in Claude Code
The Legal team and the Growth Marketing team both arrived at this independently. Use Claude.ai's conversational interface to think through your idea, explore edge cases, and generate a structured spec. Then hand that spec to Claude Code and ask it to implement one step at a time.
The Legal team adds one more instruction that helps non-technical users a lot: tell Claude Code to slow down and implement one thing at a time. The default is to output everything at once, which gets overwhelming fast.
Sub-agents for anything with multiple distinct outputs
Growth Marketing built an ad generation system with two specialized sub-agents: one focused entirely on headlines (30 character limit), one on descriptions (90 character limit). Separate agents, separate constraints, separately debuggable.
The result: hundreds of ad variations in minutes. Ad copy that used to take 2 hours now takes 15 minutes. The team went from needing dedicated engineering resources to handling it with one person.
The rule: if a workflow has two meaningfully different outputs with different constraints, it probably deserves two agents.
What non-developers actually need to know
A lawyer on the Legal team built a predictive text accessibility app for a family member with a speaking difficulty. One hour. No coding background.
The path: get an engineer to help with initial setup and permissions (this part is genuinely confusing the first time). Tell Claude your experience level in CLAUDE.md. Plan in Claude.ai first. Use screenshots instead of text descriptions. Ask for one step at a time.
The last tip from Legal is the most underrated: share your prototypes even when they're "toy" projects. The demonstrations sparked ideas across departments that wouldn't have come from documentation or presentations.
The numbers
Security Engineering cut incident resolution from 10-15 minutes to 5 minutes by feeding stack traces and documentation directly into Claude Code instead of manually scanning code.
The Inference team reduced research time by 80%. ML concepts that required an hour of reading documentation now take 10-20 minutes.
Product Design compressed a week of back-and-forth coordination on a product launch into two 30-minute sessions.
The one meta-tip
Every team that adopted Claude Code well ran internal sessions where members showed each other their actual workflows. Not documentation. Live demos. The things people invent in isolation are almost always different and usually complementary. If you're rolling this out to a team, skip the written guide and run a show-and-tell instead.
Source: How Anthropic Teams Use Claude Code — Anthropic internal report