There is an entire industry that has grown up around prompt engineering. Courses, templates, libraries, communities. The promise is that the right prompt gets you better AI output.
It is not wrong. A well-constructed prompt produces better output than a vague one.
But it is missing the point. And when I first wrote that, in May, I was missing half of one myself. More on that further down.
What is a prompt, and what does it leave out?
A prompt is a question. It tells the AI what you want right now. It does not tell the AI who you are, who your audience is, what your business does, how you speak, or what standards your content needs to meet.
Without that, the AI answers your question with the most statistically average response it can produce. That is why the output sounds like everyone else. Because without context, you are asking it to write for everyone.
What does context do that a prompt cannot?
Context tells the AI who is asking. It narrows the universe of possible answers down to the ones that fit your brand, your audience and your situation.
When your brand voice document, your audience intelligence and your business context are loaded into the session, the output changes. It is no longer writing for everyone. It is writing for your brand, to your audience, in your voice.
A mediocre prompt with excellent context produces better output than an excellent prompt with no context. Every time.
This is not just my observation. Anthropic published a piece called Effective context engineering for AI agents, and their opening line puts it plainly. Building with language models, they write, is becoming less about finding the right words and phrases for your prompts, and more about what configuration of context is most likely to produce the behaviour you want.
They are careful about one thing, and so am I. Context engineering does not replace prompt engineering. Anthropic calls it the natural progression of it. The prompt still matters. It is just no longer the part doing most of the work.
Is more context always better?
No. This is the half I missed in May, and it is the more useful half.
I wrote that you should load the brand voice document, load the audience intelligence, load the business context. Reasonable advice. It also reads like an instruction to pour everything in, and that is where it goes wrong.
Context is finite. Anthropic describe an attention budget, and every token you add spends some of it. There is a documented effect called context rot, where the more you put in the window, the less reliably the model recalls any particular thing in it. Their conclusion is worth reading twice. Context must be treated as a finite resource with diminishing marginal returns.
So the goal is not the most context. It is, in their words, the smallest possible set of high-signal tokens that gets you the outcome you want.
They also add a caution I would repeat to any business owner. Minimal does not mean short. You still have to give it enough to work with.
What happened when Anthropic cut their own instructions?
The clearest illustration came out this month. It was reported that Anthropic removed more than 80% of the system prompt behind Claude Code, and their coding evaluations showed no measurable drop.
Eighty percent of the instructions, gone, and the thing worked the same.
If that is true of the people who build the model, it is almost certainly true of the twelve-paragraph mega-prompt sitting in your notes app. Most of it is not helping. Some of it is actively crowding out the part that would.
What should you actually load?
Fewer things, chosen properly. In practice, for a small business, that is four.
- How you sound. Not adjectives. Real examples of your own writing, plus the words you never use.
- Who you are talking to. Specific enough that it rules things out.
- The facts that must never be guessed. Prices, claims, product names, the two things you are not allowed to say.
- What good looks like. One or two pieces you were happy with. Anthropic’s guidance is a small set of canonical examples rather than a long list of edge cases, and for a brand that advice holds exactly.
Notice what is not on that list. Your history, your values statement, your full service menu, three years of newsletters. Interesting to a person. Noise to a model.
How do you stop retyping it every time?
You package it. That is what a skill is, and it is the rung of the ladder almost nobody is selling. I wrote about the difference in prompt, skill or agent.
The mechanism is neat, and it is the same idea as the attention budget. A skill only loads its name and description until a task matches. Then it pulls in the detail, and only the detail it needs. Anthropic call this progressive disclosure. You get the context without paying for it on every single turn.
The same principle explains why an AI agent is a different purchase again. An agent decides its own next step, which means it also decides what to pull into context as it goes. That is more useful and considerably harder to supervise.
Where does this leave a New Zealand small business?
In a better position than most large ones, which is not the usual direction of travel.
Curating context is not a technology problem. It is a knowing-your-own-business problem. You already know how you sound, who you sell to, and what you are not allowed to claim. A twelve-person company can write that down in an afternoon. A large organisation needs six people to agree on it first.
The businesses I see getting genuinely good output are not the ones with the best prompts. They are the ones who wrote down what they already knew, kept it short, and stopped adding to it.
So the advice from May still stands, with one correction. Build the foundation before you write another prompt. Then keep it lean, because the foundation is a budget, not a cupboard.
What is in your context right now that the AI does not actually need? If you want to work through it, let’s talk.
Sources
- Anthropic, Effective context engineering for AI agents, 29 September 2025. Source of the attention budget, the smallest set of high-signal tokens, the canonical examples guidance, and progressive disclosure.
- Context rot research by Chroma, cited in the Anthropic piece above.
- Anthropic, Building Effective Agents, 19 December 2024.
- Agent Skills, open standard originally developed by Anthropic, released 18 December 2025.
- The 80% system prompt reduction in Claude Code was reported in August 2026 and is not published by Anthropic directly.
