What is context engineering, and how is it different from prompt engineering?

What context engineering is

In Effective context engineering for AI agents (29 September 2025), Anthropic separates two jobs. Prompt engineering is about writing and organizing a model's instructions. Context engineering is the set of strategies for selecting and managing everything that enters the model when it runs: the system instructions, the tools, external data, and the message history. Agents make this urgent, because each step adds tool results and messages until the useful part is buried.

The same post names the reason: "context rot", the decline in a model's ability to recall information accurately as the number of tokens in its context grows. So the advice runs against the instinct to paste everything in. Instructions should be specific enough to guide but not a pile of brittle if-then rules. A few diverse, canonical examples teach more than a long list of edge cases.

Agent Skills follow the same idea. The open specification describes progressive disclosure: an AI first sees only each skill's name and description, loads the full instructions when a task calls for them, and opens extra files only when needed. That is context engineering built into a file format.

Why a small team should care

• Everyone is already doing it, differently. Each teammate's custom instructions, pasted docs and chat history are a different context. Different context, different answers.
• Long instructions backfire. A 4,000-word system prompt built from every past complaint makes the important rules harder to follow, not easier.
• It is writing, not code. The skills are deciding what matters, cutting, and choosing good examples.

A worked example

Illustration: a made-up business. Northside Bikes, a five-person e-bike shop, has a custom assistant for customer emails. Its instructions have grown to 60 rules, and it still offers discounts it shouldn't and quotes a warranty from the old supplier.

The owner rewrites the context, not just the wording. The instructions shrink to a two-sentence role, four "can't" rules each with what to do instead ("Never offer a discount → write [Owner: discount?] in the draft"), a pointer to the one warranty document that is current, and three example replies. The old supplier's PDF is removed from the files. The next twenty drafts quote the right warranty and none offers a discount.

How to apply it this week

1. Collect the failures: five recent answers that went wrong and why.
2. Write the role and goal in two sentences at the top.
3. Turn each failure into one rule with an alternative, and delete rules no failure explains.
4. Name the source of truth per topic and remove outdated files from what the AI can see.
5. Add two or three example answers in the exact shape you want, then test the five failures again.

Common confusions

• "Bigger context windows make this unnecessary." The context rot finding says the opposite: more room is not better attention.
• "It's a new name for prompts." Prompts are one part. Documents (RAG), tools (MCP) and history are the rest.
• "Write it once." Context drifts as products and policies change. Keep it in one shared place, such as a team skill, so a fix reaches everyone.

Questions

What is the difference between prompt engineering and context engineering?

Prompt engineering is writing and organizing the instructions. Context engineering, as Anthropic describes it, covers everything that enters the model when it answers: the system instructions, tools, outside data, examples and the message history, and choosing and trimming all of it.

Why not just give the AI all the information?

Because more is not better. Anthropic describes 'context rot': a model's ability to recall information accurately declines as the context fills up. Its guidance is to find the smallest set of high-signal tokens that make the outcome you want most likely.

Is context engineering only for developers?

No. Anyone who sets up a Claude Project, a custom GPT, a skill or a support bot decides what the AI sees. The choices are the same: what instructions, which examples, which documents, and what to leave out.