Prompt engineering vs context engineering: what's the difference?

The difference

In Anthropic's words, prompt engineering refers to methods for writing and organizing an LLM's instructions for optimal outcomes. Context engineering refers to the set of strategies for curating and maintaining the right information during inference, including everything that lands in the context besides the prompt: system instructions, tools, outside data and message history. The post adds a reason the second job matters: models recall information less accurately as their context fills ("context rot"), so more input is not automatically better.

Anthropic's prompt engineering overview sets the order of work. Before you edit a prompt, have a clear definition of success, a way to test against it, and a first draft. It also notes that not every failure is a prompt problem: latency or cost can be better fixed by choosing a different model. OpenAI's model optimization guide makes a similar point: prompt engineering may be all you need for great results.

Why a small team should care

• Good prompts are team assets. The reply prompt your best support person wrote is worth keeping and sharing.
• Context is where teams differ. Two people with the same prompt and different pasted documents get different answers.
• Both need upkeep. When the policy changes, the prompt and the documents it points to must change together.

A worked example

Illustration: a made-up business. Kite Language School, an eight-person school, uses AI to draft replies to parents. The head of admissions has a prompt that works. Two teachers copy it but also paste last year's fee table into their chats, so their replies quote old fees.

The prompt was fine; the context wasn't. The school turns the prompt into a shared skill: the same instruction, three example replies, and a pointer to the current fee page instead of a pasted table. Every teacher's AI now uses the same instruction and the same source, and the fee answers match.

How to apply both

1. Write what a good result is for the task, in two or three sentences.
2. Collect five to ten real inputs to test with, including ones that went wrong.
3. Improve the prompt: clear role and goal, rules with what to do instead, and two or three examples.
4. Trim the context: point to the one current source per topic, and remove old pasted documents.
5. Save it as a shared skill so everyone uses the tested version.

Common confusions

• "Prompt engineering is dead." The prompt is still the most important part of the context.
• "Longer prompts are better prompts." Short, ordered instructions with examples beat long lists of rules.
• "Context engineering is only for developers." Anyone who sets up a project, a custom assistant or a skill chooses its context. More in context engineering, explained.

Questions

Is context engineering replacing prompt engineering?

No. Anthropic describes context engineering as the progression of prompt engineering: the prompt is still there, but the work now also covers which documents, tools, examples and history the model sees, especially for agents that run many steps.

What should I do before prompt engineering?

Anthropic's prompt engineering overview assumes three things first: a clear definition of success for your use case, a way to test against it, and a first draft prompt to improve.

How does a team keep its best prompts?

Write each one as a skill (a name, a 'use when' description and the steps) and share it from one place, so everyone's AI uses the same version and a fix reaches everyone.