Context is everything

Everyone’s obsessing over prompts. The magic words. The perfect instruction. As if the right incantation unlocks the model’s true potential.

That’s not where the leverage is.

An LLM sees one thing: its context window. Your prompt, system instructions, conversation history, tool outputs — all of it, mashed into a stream of tokens. That stream is the only reality the model knows. No memory. No persistent state. Every inference is a blank slate filled by whatever you put in front of it.

Your prompt is 50 tokens inside a 200,000-token window. A sentence on a whole page. And the whole page determines the answer.

This is why prompt engineering hits a ceiling fast. You can wordsmith all day. If the context is missing what the model needs, clever phrasing won’t save you.

The real skill is context engineering — controlling what goes into the window, when, and how.

It’s like cooking. Everyone’s fussing over technique — the flip, the timing, the presentation. But the dish is determined by what’s on the counter. Great ingredients with simple technique beats perfect technique with bad ingredients. Every time.

The ingredients are the context. Relevant code, current documentation, system state, team conventions. Get those in and the model does good work. Leave them out and it guesses — no matter how carefully you phrase the ask.

This is what tools like MCP actually do. Everyone frames them as giving AI “the ability to act.” Call APIs. Query databases. True enough. But the bigger deal is simpler: they load the right information into context at the moment it matters. You’re debugging a payment failure and the agent hits a Stripe call it doesn’t understand. It could hallucinate from training data. Or it could fetch the current API schema, drop it into context, and reason from facts instead of memory.

The model didn’t need Stripe’s entire API surface. It needed one endpoint’s schema right now. The tool fetched exactly that. Precision context, on demand.

Stop thinking about prompting. Start thinking about context. What does the model need to know? Where does that come from? What loads upfront, what gets fetched on demand?

The model is fixed. You can’t make it smarter. But you can control what it sees.

That’s the only lever you have. Use it.