Most people who use AI every day reach the same point. Rewording the question stops making much difference. What changes the answer is what the AI knows when it answers.
So they start working on that instead. They write custom instructions. They paste background into the top of chats. They keep a document of context to copy from, or connect the AI to their email and drive.
That work has a name: context engineering. If you do any of the above, you are already doing it, just by hand.
What is context engineering?
Context engineering is deciding what an AI model sees before it answers: which documents, notes, examples, history and tools go into the context window, in what order, and how much of each.
Anthropic, the company behind Claude, describes it as the natural next step after prompt engineering: the work of curating and maintaining the best possible set of information in front of the model while it answers, including everything that is not the prompt itself. Their framing of the core question is a useful one to borrow: what combination of context is most likely to get the answer you want?
A simple way to picture it: the model is a very capable new hire on their first day. Prompting is giving them instructions. Context engineering is deciding what goes in their briefing pack.
How is context engineering different from prompt engineering?
Prompt engineering improves the question. Context engineering improves the information around it. Models have become good at understanding plainly worded requests, so clever phrasing now adds less than it used to. The difference between a generic answer and one that reflects your work mostly comes down to context.
What are the parts of context engineering?
Engineers building AI products work on six things. Each has a manual equivalent you may recognise:
- Corpus: the source material. The articles, reports, videos and notes you have read and meant to keep.
- Retrieval: finding the relevant pieces. Searching bookmarks and old tabs for the thing you saved.
- Assembly: putting those pieces in front of the model in a sensible order. Copy and paste.
- Provenance: knowing where each piece came from and why it mattered. Usually kept in your head.
- Freshness: knowing when something is out of date. A flight price saved in March is not today’s price.
- Judgement: deciding what to leave out. Pasting everything skips this step.
Why does doing it by hand stop working?
Because it needs upkeep. The background document goes out of date within a couple of weeks. The bookmarks folder fills up and nobody opens it. Custom instructions get longer until they start causing problems. Each new chat means explaining yourself again and pasting the same articles again, which costs tokens and still loses detail.
Plenty of people have tried a note-taking system to solve this and given up, because maintaining it took more time than it saved.
What do you get from doing it properly?
Better answers. The AI works from your actual situation rather than general knowledge.
Fewer made-up facts. When the relevant information is in front of the model, it has less reason to invent it.
Lower cost. Sending the relevant part uses far fewer tokens than sending everything. More context is not better either: research on long contexts shows models get less reliable as more is added, and are worst at using information in the middle.
Portability. Context kept outside any one AI tool works in Claude now and in whatever you use next.
It builds up. A good prompt helps once. A well-organised body of context helps every time, and improves as you add to it.
What does this look like in practice?
Take a strategy consultant preparing for a client workshop on entering a new market.
By hand: she has read a dozen relevant reports and articles over the past month. Some are bookmarked, some are in open tabs, a few are in a Notion page. Before asking Claude to draft the workshop agenda, she spends twenty minutes finding them, pastes in four that seem most relevant, and explains the client background again. The draft is decent but misses a regulatory point from an article she forgot she had read.
With a context engineering tool: she saved each report and article into a client space in Kepler as she came across it, with a quick note on why. She asks Claude for the agenda. Claude searches that space over MCP, pulls the relevant pieces, including the regulatory article, and cites which source each point came from. Nothing is pasted, and the background she saved last month is still there for the follow-up questions.
The judgement is still hers. What changes is that finding, assembling and remembering are no longer manual. That is what we mean by context engineering for people who don’t engineer, in the same way Canva made design accessible to people who never learned Adobe. It is also the difference between memory an app recorded about you and memory you chose.
Kepler works from your browser and phone, with Claude, ChatGPT and other MCP-compatible tools. Start building your first space →
Kepler helps you create context for AI from the links you save. It improves the AI tools you already use. It isn’t one itself.
Sources
- Anthropic Engineering, Effective context engineering for AI agents (September 2025): anthropic.com/engineering/effective-context-engineering-for-ai-agents
- Chroma, Context Rot: How Increasing Input Tokens Impacts LLM Performance (2025): trychroma.com/research/context-rot
- Liu et al., Lost in the Middle: How Language Models Use Long Contexts (2023): arxiv.org/abs/2307.03172
