Prompt AI Multi-Agent Prompt Engineering
Written and maintained by KOBA42. A free original, use it in any chatbot.
A small pack of copy-paste prompts that make AI output trustworthy the way real review does: one agent builds and lists its load-bearing claims, a second is paid only to refute them, and a majority-vote variant runs several skeptics for high-stakes answers.
A single model grading its own work agrees with itself; that is not verification, it is a mirror. Splitting build from refute across separate chats removes the anchoring, and forcing the Builder to list its load-bearing claims gives the Refuter concrete targets instead of vague vibes. The N-skeptic majority vote surfaces the claims everyone quietly assumed, which is where confident wrong answers hide.
How to use it. Run the Builder, copy its answer and CLAIMS into a fresh chat, then run the Refuter. For anything high-stakes, run the Refuter three or five times in separate chats and compile with the voting prompt. Raise N for higher stakes; two skeptics is the floor.
Worked example. A model claimed a SQL migration was 'safe to run live.' The Builder listed six claims; three independent Refuters all marked 'the ALTER holds no long lock on Postgres' as WRONG for a table over a certain size, and the compiled verdict rewrote the recommendation to run it in a maintenance window.
If your team ships AI output into places where a wrong answer costs money, the assessment at koba42.com/assessment maps where these gates belong in your stack.
This is a two-role verification protocol plus a voting variant. Use it whenever an AI answer will be trusted without a human checking the work. Run the roles in separate messages or separate chats so one does not see the other's reasoning. ROLE 1, THE BUILDER. 'Answer the question below. Then, in a separate block titled CLAIMS, list every factual or logical claim your answer depends on, one per line, numbered. Do not defend them; just list them. QUESTION: [PASTE]' ROLE 2, THE REFUTER. Paste the Builder's answer and CLAIMS into a fresh chat: 'You are a skeptic paid only when you find a real error. Below is an answer and its list of load-bearing claims. For each numbered claim, return one of: SOUND, WEAK (plausible but unproven, say what evidence is missing), or WRONG (state the counterexample or contradiction). Do not comment on style. End with a verdict: does any WRONG or WEAK claim break the conclusion? If yes, say what the corrected answer would be. ANSWER AND CLAIMS: [PASTE]' THE N-SKEPTIC / MAJORITY-VOTE VARIANT. For high-stakes answers, run Role 1 once, then run Role 2 three or five times independently (new chat each time, same prompt). Then compile: 'Here are N independent skeptic reviews of the same answer. Report only claims that at least one skeptic marked WRONG, and claims a majority marked WEAK. Ignore anything all skeptics passed. Give me the shortest corrected answer consistent with the surviving reviews. REVIEWS: [PASTE ALL]' RULES. The Refuter must never rewrite for tone. A claim with no cited basis is WEAK by default, not SOUND. If the Builder and every skeptic agree, that is signal, not proof; flag anything all of them simply assumed rather than checked.
Tools used: ChatGPT, Claude, Any LLM
Want this running in your business? KOBA42 builds and operates automations like this one.