From the team behind aidrop.it — one workspace to build, host, and keep changing your code.
Prompt engineering optimized how you ask; context engineering optimizes what the model knows when you ask. As models got better, clever phrasing stopped being the bottleneck — the bottleneck moved to context: which facts, files, history, and tools are in the window when the model works. That's an engineering problem, and memory is its highest-leverage component.
The shift shows up in practice everywhere: the difference between a useless and a great answer from the same model is almost never wording anymore. It's whether the model could see the right three facts.
What context engineering is
Context engineering is the discipline of assembling the right information in the model's context window at the right time: instructions, retrieved knowledge, conversation history, tool outputs, and memory. The model is fixed; the context is the variable you control. Where prompt engineering tweaked sentences, context engineering designs the pipeline that decides what the model sees.
Why prompts stopped mattering (mostly)
Modern models infer intent well from plain language — rephrasing buys a few percent. But a model that can't see your codebase conventions or last month's decision will fail confidently no matter how artful the prompt. Garbage context, garbage output. The ceiling on output quality is set by context quality, and prompts can't compensate for missing facts.
The context stack
A working agent's context is assembled from layers: system instructions (stable rules), task input (what you asked), retrieved knowledge (facts pulled in for this task), tool outputs (what the agent observed), and memory (what's known about you, the project, the past — see AI agent memory explained). Most people hand-curate the middle layers every session — context engineering automates them.
Memory is the highest-leverage layer
Every layer except memory resets each session. Memory is the only one that compounds: each saved decision improves all future context assemblies. Context assembly becomes automatic and gets better every week you use it.
How to start engineering context
Audit a failed AI session and ask: what fact, had it been in the window, would have fixed this? Then make that fact's availability systematic — a repo file, a retrieval source, a memory entry — rather than something you paste when you remember. Repeat. Context engineering in practice is mostly turning one-off explanations into persistent, retrievable context.
FAQ
Is prompt engineering dead? No — clear instructions still matter, especially in system prompts. It's just no longer where the gains are: a mediocre prompt with great context beats a great prompt with no context.
What's the difference between context engineering and RAG? RAG is one technique inside it — the retrieval layer. Context engineering covers the whole window: instructions, history, tool outputs, memory, and the budget decisions between them. Memory stores automate the memory layer of that stack.
Do bigger context windows make context engineering obsolete? The opposite. Bigger windows fit more irrelevant text, and models still attend better to short, relevant context — selection keeps mattering more, and it gets cheaper than stuffing.
One workspace to build, host, and keep changing your code
The repository, what the project knows, the rules a change has to follow, and the path to an address — kept together, and reachable by the coding agent your team already uses.