What is CLAUDE.md and what is it for?

CLAUDE.md is a text file that Claude reads automatically at the start of each session. It contains persistent instructions: your tech stack, code conventions, style preferences, project rules. It's the way to avoid repeating the same context in every session.

When should I use /clear vs /compact?

/clear completely resets the context — like opening a new session. /compact tries to summarise the history to free up tokens without losing the thread. The Level 3 recommendation is to prefer /clear: a clean, deliberate context produces better results than a degraded summary of a long conversation.

Why does performance drop after ~100K tokens?

Language models have difficulty maintaining uniform attention across very long contexts — especially for information in the «middle» of the window. After a certain point, adding more text doesn't add more understanding: it adds noise. This degradation is why strategically resetting context improves quality.

How do I know what to include in CLAUDE.md and what to leave out?

Include only what Claude needs to know in ALL prompts for that project — not what it needs for a specific task. The criterion is: would the output change if Claude didn't know this? If the answer is yes for most tasks, it goes in. If it's only relevant for edge cases, it doesn't.

Level 3 — The Context Engineer

13 March 2026

Mindset: «You're not a coder — you're a curator»

Level 3 is where most Claude Code practitioners separate themselves from the remaining 50%. The central insight is this: output quality doesn't depend on what you ask — it depends on what information Claude has when you ask it.

Your job is no longer just to write prompts. It's to decide what goes into the context, what comes out, and when to reset. You are an information curator, not a command writer.

You're here when...

You write and maintain lean CLAUDE.md files — not bloated
You think deliberately about what Claude sees vs. what it doesn't
You use /clear between tasks to reset the context window
You monitor token usage actively
You know when less information produces better results

Skills to master

Context window management — the #1 skill at this level
Claude has a 200K token budget
Performance degrades sharply past ~100K tokens (50–60%) — the «dead zone»
Use /context to check usage at any time
Use /clear to reset (preferred over /compact)
Lean CLAUDE.md — not bloated. Less is more
Few-shot examples over long instructions
Right context at the right moment — surgical, not exhaustive

Trap: Bloated CLAUDE.md

An ETH Zurich study found that LLM-generated context files made agents WORSE in 5 of 8 settings and increased inference cost by 20%+. More context does not equal better output.

Claude needs some context, not all context. The difference between «enough» and «too much» is where quality is won or lost. A well-chosen 50-line CLAUDE.md outperforms a 500-line one accumulated without criteria. Mastering this distinction puts you in the top 50% of Claude Code practitioners globally.

How to unlock Level 4

The unlock comes when you realise you need capabilities Claude doesn't have natively. To connect Claude to your database, test your UI, or push to GitHub, you need external tools. That's Level 4.

Next steps

Context engineering is the least taught and most impactful skill in the AI ecosystem. If you want your team to develop it, we can work on it with real projects.

Frequently asked questions

What is CLAUDE.md and what is it for?: CLAUDE.md is a text file that Claude reads automatically at the start of each session. It contains persistent instructions: your tech stack, code conventions, style preferences, project rules. It's the way to avoid repeating the same context in every session.
When should I use /clear vs /compact?: /clear completely resets the context — like opening a new session. /compact tries to summarise the history to free up tokens without losing the thread. The Level 3 recommendation is to prefer /clear: a clean, deliberate context produces better results than a degraded summary of a long conversation.
Why does performance drop after ~100K tokens?: Language models have difficulty maintaining uniform attention across very long contexts — especially for information in the «middle» of the window. After a certain point, adding more text doesn't add more understanding: it adds noise. This degradation is why strategically resetting context improves quality.
How do I know what to include in CLAUDE.md and what to leave out?: Include only what Claude needs to know in ALL prompts for that project — not what it needs for a specific task. The criterion is: would the output change if Claude didn't know this? If the answer is yes for most tasks, it goes in. If it's only relevant for edge cases, it doesn't.