Grounding answers in your own sources

The most important shift in high-stakes prompting: stop asking the model what it knows, and start asking it what your documents say.

When you give the model the source material and constrain it to that, you turn it from a guesser into a research assistant whose work you can check.

The core technique is cite-or-abstain:

Answer using ONLY the sources below. After each claim, cite the source in brackets. If the sources don't contain the answer, reply exactly: 'Not found in the provided sources.' Do not use outside knowledge.

This grounds the answer, makes every claim traceable, and gives the model an honorable way to say 'I don't know.' It's the manual version of what RAG tools automate.