Quick Answer
Context engineering is choosing what information an AI sees, not just how you word the request. It covers the documents, examples and history you supply. Because model recall drops as context grows, curating what you feed in now beats rewriting the instruction.
If your AI output has stopped improving no matter how you reword the prompt, the wording is not your problem. What the model can see is. That shift has a name now, and it is the single most useful idea in prompting this year.
All 20 prompts we added to the library this week are built around it. Here is the thinking behind them.
What Is Context Engineering?
Anthropic's engineering team defines it in their guide to effective context engineering for AI agents as the set of strategies for curating and maintaining the optimal set of tokens during model inference, including everything that lands in the window outside the prompt itself.
The distinction they draw is worth holding on to. Prompt engineering is about how you write one instruction. Context engineering is about managing the entire state: system instructions, tools, external data, and the message history piling up behind you.
And it is not a one off. They describe the curation phase as happening every time you decide what to pass to the model. That reframes the job. You are not writing a magic sentence. You are deciding, repeatedly, what deserves to be in the room.
Why Do Long AI Chats Stop Working?
Because context is a finite resource, and most people treat it as free.
The same guide names the effect: context rot. As the number of tokens in the context window increases, the model's ability to accurately recall information from that context decreases. Not gradually irrelevant. Actively worse at finding what you gave it.
There is a structural reason. A transformer forms n squared pairwise relationships for n tokens, so as context length grows, the model's ability to capture those relationships gets stretched thin. Anthropic frames this as an attention budget that depletes with every token you add.
Which kills a common habit. Pasting everything you have, just in case, is not a safe default. Every irrelevant paragraph competes with the paragraph that mattered.
What Should You Actually Put In?
Aim for what Anthropic calls the right altitude. They describe a Goldilocks zone between two failure modes: prompts stuffed with brittle if-then logic for every edge case, and prompts so vague they give the model nothing concrete to act on.
In practice, four things earn their place.
- Hard constraints. Numbers, dates, names, prices, compliance requirements. Anything the model cannot infer and must not invent.
- Decisions already made. State the decision, drop the debate that produced it.
- Two or three real examples of the output you want.
- What to do when it does not know. Tell it to say so rather than fill the gap.
That last one prevents more bad output than any clever phrasing. A model with no instruction about uncertainty will produce something confident and plausible, which is the worst possible failure mode for a business document.
How Do You Cut Context Without Losing Facts?
Run everything through one test: could the model produce the right answer without this?
Background usually fails that test. So does project history, the reasoning behind settled decisions, and anything repeated. Constraints and figures always pass.
The safety net is to make the cutting visible. Ask for a list of what was removed and why, then scan it. Ten seconds of checking beats discovering that the delivery deadline got trimmed out. Our new prompts Trim a Long Brief Without Losing the Facts and Trim an AI Coding Context to What Actually Matters both work this way, and the review step is the point.
Do Examples Beat Instructions?
Usually, though relevance decides it.
Anthropic's write-up on prompting for long context tested this directly on documents of roughly 70,000 and 95,000 tokens. Generic examples drawn from unrelated general knowledge did not help performance. Examples taken from the same document did.
Pairing contextual examples with a scratchpad for pulling relevant quotes before answering moved accuracy from 0.939 to 0.961, described in the write-up as a 36 percent reduction in errors. Worth being precise about what that number is: it comes from the models available at the time of that testing, so treat the technique as durable and the figure as a snapshot rather than a promise.
The practical version is simple. Stop writing paragraphs describing your tone. Paste three things you actually wrote and say match this.
When Should You Start a Fresh Chat?
Sooner than feels natural. The signals are consistent: answers drift, the model contradicts something it agreed 40 messages ago, or you find yourself repeating a constraint you already gave.
That is context rot doing exactly what the research describes. Rewording your next message will not fix it.
The move is a carry forward note holding decisions, constraints, open questions, and any wording that must survive exactly. Nothing else. Not a recap of the conversation, not your original brief again. Shorter genuinely beats complete here, which is the opposite of what most people assume.
How Does This Apply to Teams in Asia?
Context is where local knowledge lives, and it is the part a general model does not have.
A model can write a competent product description. It cannot know that your peak sale dates are 9.9 and 12.12, that your buyers expect a particular payment method, or which language belongs on which channel. None of that is in its training in any reliable, current form. It has to come from you.
This is also where the cost of guessing is highest. Invoicing rules, licensing claims and pricing display requirements differ across the region and change. Put the verified facts in your context pack, mark the unverified ones, and check those against the official regulator before anything goes out. Our guide to AI prompts for small business owners covers the everyday jobs, and prompts for entrepreneurs goes wider on planning.
Build the pack once, reuse it all year, and refresh the parts that go stale. That is most of the benefit for a small team.
What Else Do People Ask?
What is context engineering?
Anthropic defines it as the set of strategies for curating and maintaining the optimal set of tokens during model inference, covering everything that lands in the context window rather than just the prompt you type. Prompt engineering is about wording one instruction. Context engineering is about managing the whole state, including documents, examples, tools and message history.
Why do long AI chats get worse over time?
Anthropic calls it context rot: as the number of tokens in the context window increases, the model's ability to accurately recall information from that context decreases. The architectural reason is that a transformer forms n squared pairwise relationships for n tokens, so attention gets stretched thinner as the conversation grows. More context is not automatically better context.
Do examples work better than instructions?
Usually yes, but only relevant ones. In Anthropic's long-context tests, generic examples on general knowledge did not help performance, while examples drawn from the same document did. The practical rule is to show two or three real samples of the output you want rather than writing another paragraph describing it.
How much does better context actually improve accuracy?
In Anthropic's published long-context experiments on documents of roughly 70,000 and 95,000 tokens, combining a quote-extraction scratchpad with contextual examples moved accuracy from 0.939 to 0.961, which the write-up describes as a 36 percent reduction in errors. Those figures come from the models of that time, so treat the technique as durable and the exact number as a snapshot.
When should you start a fresh AI chat?
When answers start drifting, contradicting earlier ones, or forgetting decisions you already made. That is context rot rather than a bad day for the model. Write a short carry forward note containing decisions, constraints and open questions, then paste it into a new chat. Keeping the summary short usually beats making it complete.
Sources: Anthropic engineering, Effective context engineering for AI agents, for the definition of context engineering, the context rot finding, the attention budget framing, the n squared pairwise relationships explanation, and the right altitude guidance. Anthropic, Prompt engineering for Claude's long context window, for the tested techniques and the accuracy movement from 0.939 to 0.961 across roughly 70,000 and 95,000 token documents. Accuracy figures reflect the models tested at the time of that research. Verified August 2026.