Now in the Chrome Web Store

Read at the depth you have time for.

One slider on any article. Slide left for a glance, right for a Socratic deep dive — and everything in between. Same page, five reading depths.

Add to Chrome — Free

Works on any article No account needed

claude.com/blog/how-claude-code-works-in-large-codebases

Enterprise AI · Claude Code at scale

How Claude Code works in large codebases: Best practices and where to start

Anthropic · May 14, 2026 · 5 min read

Claude Code is running in production across multi-million-line monorepos, decades-old legacy systems, and distributed architectures spanning dozens of repositories. These environments present challenges that smaller, simpler codebases don't.

This article covers the patterns we've observed that have led to successful adoption of Claude Code at scale. We use "large codebase" to refer to a wide range of deployments: monorepos with millions of lines, legacy systems built over decades, dozens of microservices…

Depth

How Claude Code works in large…

2,140 WORDS · 5 MIN

Reading depth 01Glance Summary Read Quiz Deep Dive

01Glance

02Summary

03Read

04Quiz

05Dive

Glance ~ 8 sec read

The harness around the model — CLAUDE.md, hooks, skills, plugins, and MCP servers — shapes Claude Code at scale far more than the model alone.

Summary ~ 45 sec · 7 points

01Claude Code navigates like a human engineer — files, grep, references — not a stale embedding index.
02The harness matters as much as the model.
03Keep CLAUDE.md lean and layered — root for the big picture, subfolders for local detail.
04LSP integrations give symbol-level navigation in multi-language repos.
05Hooks are the highest-leverage extension point — a stop-hook can propose CLAUDE.md updates while context is fresh.
06Use MCP servers to connect Claude to internal tools, dashboards, and APIs it can't otherwise reach.
07Smooth rollouts assign a DRI and revisit config every 3–6 months as models evolve.

Read ~ 4 min · simplified

Grade 10 · Trimmed −68% · Key terms 6

How Claude Code finds its way

Claude Code navigates a codebase the way an engineer does — reading files, using grep, following references. It avoids the failure mode of RAG-based tools whose embeddings go stale faster than active teams can refresh them.

The tradeoff: agentic search works best when Claude has enough starting context to know where to look. That context comes from CLAUDE.md files layered through the tree.

The harness, not the model

The harness is built from five extension points: CLAUDE.md, hooks, skills, plugins, and MCP servers. The order in which teams build them matters — each layer builds on the one before it.

For multi-language repos, LSP integrations are one of the highest-value investments: they give Claude "go to definition" and "find all references" precision.

Where to start

Assign a single DRI. Get the root CLAUDE.md lean and accurate first. Then layer in hooks and skills as your team's patterns emerge.

Quiz Question 1 of 4

Q1What does "the harness" refer to in this article?

AA new training method

BThe ecosystem around the model

CA safety constraint

Deep Dive Socratic · live

If a team picked the most capable model but skipped CLAUDE.md, hooks, and skills — what would go wrong first?

It'd waste tokens hunting for context every session.

01Glance

02Summary

03Read

04Quiz

05Dive

Five depths. One slider. Pick the one that fits the next two minutes.

Every depth is generated up front from the same article. Drag the slider — no reloads, no waiting.

018 sec

Glance

The whole article distilled into one sharp sentence. Decide whether to keep going.

0230 sec

Summary

Four to six bullets in the author's voice. The shape of the argument, not a bag of facts.

034 min

Read

A simplified version of the full piece with key terms inline. Hover anything to learn it.

043 min

Quiz

Four questions on what actually matters. Wrong answers go into spaced repetition.

05live

Dive

A Socratic counter-question that pushes you past the article — and answers back.

A note from the creator

Hey, I’m Wenyu.

English isn’t my first language, and long paragraphs used to make me feel tired before I even started reading. I built Depth to make reading feel more gradual — start simple, go deeper when you’re ready, and absorb the idea at your own pace.

Hope you like it.

Meme: reading a wall of text vs. reading through Depth's progressive layers — — how it feels, basically