From file to context

Having the knowledge written down is not enough. It needs to reach the agent at the right time, and in pieces.

This is where three mechanisms come in that do different things and are constantly confused.

RAG retrieves what is relevant and injects it before the model responds. This is how the company's knowledge base becomes input without retraining any model. Watch the trap: loose semantic search over a large repository pulls the wrong document with convincing text. Filter before searching.

Skills do the opposite of RAG: instead of retrieving information, they load procedure. A folder with instructions, approved examples, and a validator, which enters the context only when the task is of that type.

Context engineering is the discipline that decides what goes in, when, and in what format. It does for context what capital management does for cash, and separates the agent that knows what matters from the one that received everything and prioritized nothing.