Context Packs

Context Packs are the foundation of reliable AI systems. You will use them every time you build an agent, a retrieval-augmented generation (RAG) pipeline, or any automated task that needs specific knowledge. Mastering this concept lets you move on to Prompt Iteration and Prompt Optimization — without a well-assembled Context Pack, those techniques have nothing to work on. It also builds directly toward Context Engineering, where you will decide what information to include, in what order, and how to keep it fresh.

Imagine you are a new employee at a company. On your first day, your manager hands you a thick folder. Inside is a list of your responsibilities, examples of how previous employees solved similar problems, and a note about the company's style — be formal in emails to clients, use bullet points in reports, ask before changing anything in the shared database. You read the folder, and now you know how to work without asking your manager every five minutes. A Context Pack is exactly that folder, but for an AI model. It is a bundle of text that you give to the model before it starts answering. The text contains everything the model needs to do the task well: the instructions (what to do), the examples (how to do it), and the conventions (the rules to follow). Without the pack, the model is like a new employee with no folder — it guesses, makes mistakes, and does things in a random way. With the pack, it acts like a seasoned professional, because you have given it the same knowledge a human would get from experience. The trick is that the model does not remember anything between requests — every time you call it, it is a fresh employee. So you must give it the folder every single time. That is why you build a Context Pack once and reuse it.

Think of a Context Pack as a recipe card for a complex dish, say a traditional Hyderabadi biryani. The recipe card has three parts: the list of ingredients (the specifications), a picture of the finished dish (the example), and notes like 'always use aged basmati rice' or 'do not open the lid while cooking' (the conventions). If you hand this card to a chef, they can cook the dish without asking you questions. If you hand it to ten different chefs, they will all produce something recognizable as biryani, though each will be slightly different. Now imagine you did not give the card — you just said 'cook biryani.' Each chef would do something wildly different: one might add tomatoes, another might use instant rice, a third might make it too spicy. The recipe card is the Context Pack: it turns a vague request into a specific, repeatable result. But the analogy breaks down in one important way: a human chef can remember the recipe after cooking it once, and they can adjust if they run out of an ingredient. A language model cannot remember anything between calls. It only sees the card you give it right now. If you forget to include the note about aging the rice, it will happily use regular rice. So the Context Pack must be complete every time — you cannot assume the model picked anything up from a previous conversation.

A Context Pack is a deliberately assembled bundle of text — containing specifications, examples, and conventions — that you give to an AI model before it performs a task, to provide all the context it needs in one reusable, reviewable package.

You already know that a prompt is the text you give a model. A Context Pack is a special kind of prompt — one that is designed to be reused and shared. It is a way of turning the knowledge that lives in your head or in your organisation's documentation into something the model can use. This concept connects to Prompt Iteration because you will refine your Context Pack based on test results, and to Prompt Optimization because you can treat the pack's contents as a set of choices to be improved automatically. It also connects to Structured Output Constraints: a good Context Pack often includes a schema that forces the model to reply in a specific format. And it is the foundation of AI Agents, because an agent needs a pack that describes its tools, its goals, and the rules for using them.