Why context is (almost) everything

There's a single principle underneath all of context engineering: a model's output is only as good as the context it's given. Get the context right and even a simple request succeeds; get it wrong — missing key information, or drowned in irrelevant noise — and even a beautifully worded prompt fails. This lesson makes that principle concrete.

Garbage in, garbage out — but also 'not-enough in, wrong out.' A model can only reason over what's in front of it. If the fact it needs isn't in the context, it can't use that fact — it will either say it doesn't know or, worse, guess. And if the context is cluttered with irrelevant material, the model's attention gets diluted and quality drops (Module 3 shows exactly how). So two failure modes bracket the goal: too little of the right information, and too much of the wrong. Context engineering is the practice of hitting the target between them.

The goal: the smallest set of high-signal tokens. Anthropic captures the aim precisely: good context engineering means finding the smallest possible set of high-signal tokens that maximize the likelihood of your desired outcome. Notice both halves. High-signal — the information genuinely has to be relevant and useful. Smallest possible — because every token you add competes for the model's finite attention (Module 1's next lesson). More is not better; more of the right thing, and less of everything else is better. This single sentence is the north star of the whole course.

Why a perfect prompt isn't enough. People often obsess over prompt wording while ignoring context, and then wonder why results are inconsistent. But a flawless question with the wrong background still fails; a plain question with exactly the right context succeeds. That's why the field's center of gravity moved from "prompting" to "context": once you accept that the model can only work with what it's given, assembling what it's given becomes the main lever — and phrasing becomes one part of a larger craft.

Not 'prompt engineering is dead.' To be clear and honest: this doesn't mean wording stops mattering. A clear instruction is still high-signal context. The point is that wording is one ingredient, and the bigger wins usually come from getting the right information into the window and keeping the wrong information out. Context engineering elevates prompting rather than discarding it.

The mindset: a model's output is only as good as the context it's given — it can only reason over what's actually in front of it, so missing information yields ignorance-or-guessing and cluttered information dilutes attention and drops quality. The goal, in Anthropic's words, is the smallest possible set of high-signal tokens that maximize the desired outcome: relevant enough to help, small enough not to drown the model. A perfect prompt with the wrong context still fails, which is why assembling the right context is the main lever — not a replacement for good wording, but the larger craft that good wording lives inside.