Provide additional context

Providing context means giving the model the background information it needs but doesn’t have: your codebase conventions, the customer’s plan tier, the document being discussed, what was already tried. An LLM only knows two things — its training data and what’s in the current prompt. Anything about your situation that isn’t in the prompt effectively doesn’t exist, so the model will improvise, and improvisation is where hallucinations come from.

For software builders this is the difference between a demo and a product. A generic model answering “How do I reset my password?” invents a plausible flow; the same model given your actual help-center article answers correctly. Grounding responses in supplied context is the core idea behind retrieval-augmented generation, and it starts here: reliable AI systems are mostly systems that assemble the right context at the right time.

In practice, you’ll build prompts that inject context programmatically: fetch the user’s recent orders, the relevant docs snippet, or the current file contents, then wrap them in clear delimiters like <context>...</context> above the instruction. Pair the data with a grounding rule — Answer using only the context above; if it's not covered, say so — and be selective: context windows are large in 2026, but irrelevant filler still dilutes attention and inflates cost. Curate what you include; don’t just dump everything.