Phase 2: Context Engineering for Builders · 40 min · Claude Code · Python

What Context Engineering Actually Is

The context window is RAM, not a diary. What you load, and what you evict, is the actual engineering.

Hiring signal: Explaining why 'just paste more context' makes an agent worse, not better, shows you understand context engineering as an active discipline rather than a synonym for 'give it more information.'

What you will learn

Explain the context window as a RAM-like budget: small, fast, expensive, and volatile
Name the four context types: instructions, knowledge, tools, feedback
Explain why adding more context can degrade an agent's output instead of improving it
Classify a real context item into its correct type and compute whether a context budget has been exceeded

Introduction

What Context Engineering Actually Is

An engineer's agent session starts strong: clean edits, correct reasoning, fast turnaround. Three hours in, on the same long session, it starts making decisions that don't track — suggesting a fix already tried and rejected twenty minutes ago, misreading which file it's currently editing, contradicting an instruction from earlier in the conversation. His instinct is to paste in more: the full error log, an entire related file "just in case," a summary of the whole conversation so far pasted back into itself. Each addition makes it slightly worse, not better. He's treating the context window like a diary he can keep adding pages to. It doesn't work like a diary. It works like RAM.

The context window is RAM, not storage

A widely-cited framing of this makes the mechanism precise: "The context window is RAM. It is small, fast, expensive, and volatile. Everything the model 'knows' in this turn — the system prompt, the conversation, retrieved documents, tool results, memory — has to be resident in that RAM at inference time. Nothing else exists to the model." Storage (a wiki, a database, a vector store) can hold effectively unlimited information because nothing has to actively process all of it at once. RAM is different: everything loaded into it competes for the same fixed, finite space, and — critically for a transformer — competes for the model's attention across all of it simultaneously. Adding more to a full context window isn't like adding more pages to a notebook. It's like running more programs on a machine that's already near its memory limit: things that used to work start behaving strangely, and it's not obvious from the outside which addition caused it.

Why "paste more context" backfires
Every token in the context window is competing for the model's attention with every other token. Irrelevant or redundant context doesn't just take up space — it actively dilutes the model's focus on what actually matters for the current step. Pasting an entire file "just in case" when only one function in it is relevant adds noise the model has to process alongside the signal, and more noise measurably degrades performance on the actual task, even though it feels like you're being helpful by providing more.

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers The four context types, Naming the type before deciding what to do with it, Build It — plus a hands-on lab, quiz, and project artifact.

Create a free account to unlock Phase 0 and Phase 1 of every course — no credit card.