L4Free

Context Engineering Fundamentals

Your context window is RAM. Treat it like it.

After this, you'll be able to identify the three context layers, prune stale history before it degrades output, and use /clear as a phase boundary instead of a panic button.

Before you start

Complete Thinking in Agentic Sessions first; this lesson builds on the CLAUDE.md and session-structuring habits you established there.

The idea

Context engineering is the discipline of deciding what goes into the model's working memory, and what stays out. Most people treat the context window like a clipboard. Paste in everything. Hope for the best. This is why long sessions degrade.

A model with 200K tokens doesn't mean you should fill 200K tokens. Signal-to-noise in context is the real variable. 10K tokens of sharp, relevant context beats 80K tokens of everything you could think of.

Here is the before and after: a session with 80K tokens of full file pastes missed a refactor opportunity the model would have caught in a focused 8K session. The 200K spec used less than 4% of effective attention before saturation. Pruning to 8K of relevant context (the broken function, its types, and two related call sites) produced the correct fix in two exchanges instead of twelve.

The three context layers every L4 engineer manages deliberately: (1) System prompt, your permanent instructions and persona. (2) Working context, the current task, files, and constraints. (3) History, previous turns. Most people never prune any of these. L4 engineers treat stale context like a memory leak.

The /clear command in Claude Code is not for when things break. It's for when you finish a phase. Treat it like a commit, you checkpoint, then clear. Fresh context per work unit. CLAUDE.md is the bridge between sessions, your persistent instruction layer that survives /clear.

Try it (5 min)

Watch out for

Treating /clear as a panic button after things break. By that point you have already lost the diagnostic context you would need to debug.
Using context-window size as a goal. 200K available does not mean 200K should be filled. Signal-to-noise is the real variable.
Forgetting that CLAUDE.md counts against your token budget every session. Bloated CLAUDE.md is a permanent tax.
Pasting whole files when one function or one section is what you actually need. Whole-file pastes are the most common form of token waste.
Letting one mega-session run for 80+ turns. Past about 30-40 turns, coherence drops well before the token ceiling does.

Paste this into Claude:

Open your longest active Claude session today. Paste a rough description of what's in it: how many turns, what topics have come up, which files or large pastes are still in scope. Then ask: 'Audit this session. Which turns are stale, meaning they no longer relate to what I'm doing now? Which file pastes could be replaced with a 5-line summary? What is load-bearing and must stay?' Use the answer to start a fresh session with only the load-bearing content plus a one-sentence framing.

What good looks like:

You identified at least three turns or pastes in your current session that are stale and contributing zero signal
You restarted with a clean context that contains only load-bearing items plus a one-sentence framing
The first response in the fresh session was more focused than the last few responses in the original

When this breaks

Breaks when you expect more context to fix a degrading session because the degradation is caused by accumulated noise, not by missing information.
Breaks when stale system prompt entries contradict newer working context because the model cannot tell you which version is authoritative and silently picks one.
Breaks on tasks that genuinely require the full long history (multi-day investigations, large narrative drafts) because aggressive pruning removes connections that matter.

You can now

Identify the three context layers in your current session and name one specific item in each that should be pruned before your next request.

Key takeaways

Context is RAM, not a clipboard. Signal beats size, fresh context per work unit beats one giant session, and pruning is a habit, not an emergency move.

Context window equals RAM, not a clipboard. Signal-to-noise is the real variable.
10K tokens of sharp relevant context beats 80K tokens of everything you could think of.
Prune stale history like you prune dead code. Treat it as a memory leak, not as backup.
/clear is a phase boundary, not a panic button. Checkpoint first, then clear, then resume.
CLAUDE.md is your persistent context layer. Everything else resets every session.

Open your longest active Claude session today. Paste a rough description of what's in it: how many turns, what topics have come up, which files or large pastes are still in scope. Then ask: 'Audit this session. Which turns are stale, meaning they no longer relate to what I'm doing now? Which file pastes could be replaced with a 5-line summary? What is load-bearing and must stay?' Use the answer to start a fresh session with only the load-bearing content plus a one-sentence framing.