4 prompts
Most hallucinations are preventable before you ask anything. How you load and instruct the notebook matters more than any single question.
Set once per notebook. Every answer arrives pre-wired for verification, and “not in the sources” replaces confident guessing.
Every claim must carry a citation. Name the source document and the section or page for each substantive statement. If sources conflict, cite both sides. If the answer is not in the sources, say exactly: “Not in the sources.”
Generate a one-page glossary of your project's acronyms and names, then add the output back as its own source. Cross-references stop misfiring.
Create a one-page glossary of the acronyms, project names and people mentioned across these sources: term — expansion — one-line role. I will save this as its own source so future answers resolve these references correctly.
Belt-and-braces with the source checkboxes: names the allowed sources inside the prompt so the answer can't quietly borrow from the rest.
Answer using only [source names]. Ignore all other sources in this notebook, even where they seem relevant. If the selected sources cannot answer the question, say what is missing instead of borrowing from the others.
Nudges the model's defaults from quick-and-shallow to long-and-enumerated for the whole notebook.
Default to thorough over fast. Prefer long, complete answers that work through the sources section by section. Never compress a list of items into “several” or “various” — enumerate them.