The Claude Code Setup Behind Anthropic’s Own Engineers (Exact Config You Can Copy)

The people who built Claude Code spend over $1,000/month using it. They run 5 instances in parallel and ship full features without typing a single line of code.

Their entire setup is one folder you already have on your machine. You just never configured it.

Here’s their exact config, copy it in 5 minutes 👇

🧠Before we dive in, I share daily notes on AI & vibe coding in my Telegram channel: https://t.me/zodchixquant

Pattern 1: CLAUDE.md as an engineering manager

Boris Cherny’s CLAUDE.md reads less like a config file and more like expectations from a senior engineering manager. Not “use TypeScript” but “here’s how our team makes decisions.”

Key insight from Anthropic’s best practices docs: every line should answer “would removing this cause Claude to make a mistake?” If the answer is no, delete it.

A 5,000-token CLAUDE.md costs 5,000 tokens every turn.

Pattern 2: The interview pattern

Before building any large feature, Anthropic engineers have Claude interview them first using the AskUserQuestion tool:

After the spec is written, they start a fresh session to execute it. Fresh context = no bias from the planning conversation.

This consistently produces better results than going straight to coding because Claude catches edge cases during the interview that you wouldn’t think to mention in a prompt.

Pattern 3: Writer/Reviewer separation

Anthropic engineers never let Claude review code it just wrote. The same session that wrote the code is biased toward defending it.

Instead, they use two separate sessions:

Same pattern works for tests: one Claude writes tests, another writes code to pass them. The adversarial setup catches bugs that self-review misses.

Boris Cherny takes this further with subagent fan-out: spawn several agents checking style, bugs, and history in parallel.

Then spawn 5 more agents whose job is to challenge the first group’s findings. The result filters out false positives and catches real issues.

Pattern 4: Skills as folders, not files

The most common misconception about skills: they’re “just Markdown files.” Internally, Anthropic treats them as full folders with scripts, assets, data, and reference docs.

An Anthropic engineer built the frontend-design skill iteratively with users to improve Claude’s design taste and avoid clichés like Inter font and purple gradients.

The most valuable section in any skill is the Gotchas section: common mistakes based on real problems Claude encounters.

Internally, their skills fall into categories:

Verification skills are the highest-value: they include scripts inside the skill, record video of Claude’s output for visual verification, and add programmatic assertions at every step.

Anthropic says it’s often worth dedicating an engineer for a full week to polish one.

Pattern 5: Plugins from the marketplace

New in 2026: plugins bundle skills, hooks, subagents, and MCP servers into a single installable unit.

Browse and install community and Anthropic-built plugins. One command adds an entire capability.

If you work with a typed language, the code intelligence plugin gives Claude precise symbol navigation and automatic error detection after edits.

This is how Anthropic scales internally: instead of every engineer configuring their own skills from scratch, shared plugins standardize the setup across teams.

Pattern 6: Hooks for automated quality

Anthropic’s internal hooks run automatically at every stage of a session:

Every file auto-formats after Claude writes it. Every log file gets filtered before Claude reads it. No manual steps and forgotten formatting or wasted tokens on 10,000-line logs.

Pattern 7: Batch operations with headless mode

For large migrations or analyses, Anthropic engineers distribute work across many parallel Claude invocations:

Each invocation runs in its own context with scoped permissions and a budget cap. No single task can run away. The engineer monitors progress and reviews results.

The full settings.json (copy-paste ready)

The workflow they actually follow

Every complex task at Anthropic follows this loop:

Step 7 is the most underrated. “Update CLAUDE.md so this doesn’t happen again” is the most powerful prompt in Claude Code.

Over time, your CLAUDE.md captures every mistake Claude has made in your project and prevents all of them automatically.

Thanks for reading!

I share daily notes on AI, finance, and vibe coding in my Telegram channel:

https://t.me/zodchixquant