Practice12 min read

Source-Grounded Prompting

The hallucination antidote: don't ask AI to remember — give it the source

If you've ever caught AI inventing a citation, you've experienced the core failure: AI's "memory" is unreliable for specifics. The fix is simple. Stop asking it to recall facts. Start giving it the source material and asking it to work from that.

This is called source-grounded prompting. It changes AI from a flaky encyclopedia into a reliable analyst.

The source-grounded pattern

Three ingredients:

Paste the source — full text, a relevant excerpt, or summarized notes. The more direct the source, the better.

Explicitly constrain the answer — "Base your answer ONLY on the source I just gave you. Do not bring in outside information."

Demand acknowledgment of gaps — "If the source doesn't answer my question, say 'not in the source' rather than inferring."

This pattern works for: summarizing a long report, pulling key data from meeting notes, answering questions from a contract, synthesizing customer feedback, comparing claims across multiple documents.

Paste, don't summarize first. Some people pre-summarize the source before pasting, thinking it'll save AI work. Don't. The summary step is exactly where errors creep in. Let AI work from the actual source — that's the whole point.

Practice Exercise

Your team just finished a quarterly customer feedback round — 47 written responses from B2B customers about your product. Your manager wants a clear summary by tomorrow: what are the top 3 themes, with representative quotes for each? You're worried about AI inventing themes that aren't really there, or attributing quotes to themes that don't fit. Write the prompt that handles this carefully, given you'll paste all 47 responses into it.

No pressure — just give it your best shot! Write a prompt for the scenario above and our AI will give you friendly, specific feedback on how to improve.

0 / 5,000

(try writing your own prompt first!)