Leaked NVIDIA prompt engineering playbook reportedly cuts hallucinations by 41%

On October 22, X user @thisguyknowsai posted that an internal AI prompt handbook from NVIDIA engineers had leaked, and broke down its six anti-hallucination prompt templates one by one in a thread, relaying internal test data: a single "refusal protocol" prompt cut hallucinations by 41%, and source-interrogation prompts reduced fabricated citations by 67%.

Confirmed

• The six templates form a complete anti-hallucination workflow: before answering, "boundary setting" (listing forbidden behaviors — no fabricating unlabeled statistics, no naming real people when uncertain, no presenting opinions as facts — framed as hard stops, not guidelines) and "assumption audit" (the model first lists all its assumptions about what the user knows, the output's purpose, constraints, etc., and asks the user to confirm each before starting); during, a "refusal protocol" (write UNCERTAIN: [what is unknown] when facts are uncertain, CANNOT COMPLETE: [reason] when the task can't be done, never fill gaps with assumptions); after, "confidence scoring" (tag each factual claim HIGH/MEDIUM/LOW so users only verify LOW items), "source interrogation" (for each key claim, note the source type, time range, and confidence level; mark FLAG if unavailable), and "red-team review" (paste the output back and have the model critique it as a hostile reviewer, flag issues with [FLAG], then produce a fixed clean version).
• The thread author is @thisguyknowsai, and all six posts were published on October 22.

Unconfirmed

• Whether the handbook is genuinely an internal NVIDIA document, and the provenance of figures like 41% and 67%, rest solely on the poster's account; NVIDIA has not confirmed them.

Why it matters

• Authentic or not, the six templates are directly usable and cover the full pre-/in-/post-task hallucination-control pipeline. The red-team review ("let the model self-critique then fix") and the assumption audit ("align assumptions before working") are especially practical references for workflows that rely heavily on LLM output.

2026-10-22 ~ 2026-10-22 · 7 related posts

• [source] Leaked NVIDIA internal prompting playbook claims 41% hallucination cut — thisguyknowsai · 2026-10-22
• Boundary Setter prompt: hard constraints over soft instructions — thisguyknowsai · 2026-10-22
• [source] Refusal Protocol prompt: make models admit UNCERTAIN to cut hallucinations 41% — thisguyknowsai · 2026-10-22
• Confidence Scorer prompt: make models label every claim HIGH/MEDIUM/LOW — thisguyknowsai · 2026-10-22
• [source] Source Challenger prompt reportedly cuts fabricated citations by 67% — thisguyknowsai · 2026-10-22
• Assumption Audit prompt: list assumptions first, then confirm before answering — thisguyknowsai · 2026-10-22
• Red Team Pass prompt: make the model critique and fix its own output — thisguyknowsai · 2026-10-22