Knowledge
From prompt engineering to context engineering
The right word in your prompt matters less than the right information in the model's working memory.

A prompt is the question you ask. The context is everything the model sees when it answers: your question, the rules, examples, documents, tools and what happened before.
Plain technical English Context is the set of tokens that the model reads before it answers. More tokens make it harder to recall one fact. So put only what the next step needs in the context.
Why it matters
A good answer depends more on the context than on the words in your question. Anthropic describes context rot: the more tokens in the context window, the worse the model recalls information from it. Context is a limited resource.
The term context engineering took off in June 2025. Simon Willison quotes Andrej Karpathy: it is about “filling the context window with just the right information for the next step”.

How it works
The context has five parts; only what belongs to this task goes in.
The Claude Code guide says the same: keep the instruction file short, put team procedures in skills, and give the model a check to run.

What you do
- Write the task in one sentence, with how you see that it is done.
- Put fixed rules in one short file that the model reads every time.
- Put ways of working in separate skills, which the model loads only when it needs them.
- Add only the documents that belong to this task.
- Add a test or check that the model can run by itself.
At Denkbaar, a task for an agent stands on its own: where the work is, the goal, what is not allowed, and when it is done. The agent cannot ask back. So everything it needs is in it, and nothing more.
What we can do with you: we prepare your knowledge and rules so that an AI assistant works well with them every time. Book a conversation.