The two definitions side by side
Anthropic’s engineering team drew the line in September 2025. Prompt engineering is “methods for writing and organizing LLM instructions for optimal outcomes”. Context engineering is “the set of strategies for curating and maintaining the optimal set of tokens” during inference, which includes everything that lands in the window, not only the instructions.
| Aspect | Prompt engineering | Context engineering |
|---|---|---|
| Unit of work | One instruction or system prompt | Everything in the window, turn after turn |
| Time span | Instructions reused across one or many calls | Context assembled for each call and maintained across turns |
| Main question | How do I phrase this? | What should the model see right now, and what not? |
| Typical tools | Wording, examples, role, format | Retrieval, tool design, compaction, notes, sub-agents, memory |
| Typical failure | Vague or contradictory instruction | Right instruction drowned in irrelevant history or stale data |
| Who owns it | The person typing | Whoever builds or configures the agent, and the team that feeds it |
Why the shift happened
A chat answer is one call: you write a prompt, the model answers. An agent loops. It calls a tool, reads the result, calls another tool, and every result is added to the window. After an hour of work the original prompt might be a few hundred tokens in a window of hundreds of thousands. Whether the agent stays on track then depends far more on what else is in there than on how well the first instruction was phrased.
That is why the techniques in context engineering are about flow rather than wording: load files only when needed, clear old tool output, summarize history when it gets long, write notes to a file outside the window, and hand side tasks to a sub-agent with its own clean window. The Manus team adds a practical one: keep the start of the context stable so the model provider’s cache can be reused, because rewriting the top of the context on every turn costs both speed and money.
The common misunderstanding
The mistake is treating the two as competitors, or treating context engineering as “a longer prompt”. Pasting your whole handbook into the system prompt is a context choice too, just a poor one: it buries the instructions that matter under material the task does not need. Claude Code’s documentation is explicit: keep each CLAUDE.md under about 200 lines, because longer files use more context and are followed less reliably, and contradicting rules may be applied arbitrarily.
An example makes the line visible. A support agent answers a return question wrongly. If it has the right facts but the wrong format or tone, that is a prompt problem: improve the instructions. If it follows the instructions but quotes last year’s return policy, that is a context problem: inspect what was retrieved and why the old version was in there.
Good context engineering often makes the prompt shorter. The handbook moves out of the prompt into a place the agent can search, and only the part that matters for this task comes in.
What changes for a team
When one person writes prompts, the knowledge lives in their head and their chat history. When a team does context engineering, the knowledge has to live somewhere shared: instruction files in the repo, a team knowledge base, rules that apply in certain situations. The prompt is still personal. The context becomes shared infrastructure.
That raises questions a prompt never had: who updates it, how do you keep it from contradicting itself, and how does what one colleague corrected reach the AI of the next. See how to set up context engineering for a team for a practical setup, and what is context engineering for the full list of techniques.
Where Bombos fits
Bombos sits on the context side of this line. It is a plugin that gives every AI tool a team uses, from Claude and ChatGPT to Copilot, Cursor and your own agents, one shared memory with team knowledge, learned rules and what already lives in mail, documents and repos. Your people keep writing their own prompts; the context behind them is the same for everyone, and you can trace what an answer was based on. It runs in the cloud or on your own server, and one person can start with their own agents.
Related questions
Is prompt engineering dead?
No. The system prompt is still part of the context and still needs care. What changed is that for agents the prompt is a small share of what the model sees, so it is no longer the only lever.
Which one should a team learn first?
Both are quick to learn at a basic level. For a team, the bigger gain is usually in context: shared instructions, shared knowledge and a habit of keeping sessions focused.
Is a CLAUDE.md file prompt engineering or context engineering?
It is written like a prompt, but deciding what goes in it, how long it is, and when it loads is context engineering. Claude Code loads it at the start of every session.
