12 debate prompts for ChatGPT and other AI assistants
Copy these original prompts to rehearse objections, switch sides, improve openings and check missing evidence. Each prompt has a specific task and a clear way to assess the output.
4 min readThe debagle team
Start with a bounded exercise
These are practice instructions you can adapt for an AI assistant. They do not assume a particular paid plan, browsing feature or ability to verify sources. Use a hypothetical issue or public information; omit private details about other people.
Replace “our town” with the actual scope if you have permission to discuss it. After each exercise, read the response as a suggestion. An articulate criticism can still misunderstand your argument.
Prompts for opening arguments
- Claim check: “Here is my opening: [text]. Identify the policy I propose, who acts and the time frame. Ask a clarifying question if any of these is missing.”
- Reasoning bridge: “Separate my claim, evidence and inference. Show where the inference needs an extra premise. Do not add new facts.”
- Ninety-second opening: “Help me cut this opening to one proposal, two reasons and one acknowledged tradeoff. Preserve uncertainty and do not strengthen factual claims.”
- Concrete stakes: “List who might benefit and who might bear costs from this hypothetical proposal. Mark these as possibilities, not verified outcomes.”
Prompts for responding to objections
- One objection: “Challenge this position with one substantive objection. Wait for my answer before continuing. Avoid personal attacks and invented studies.”
- Direct answer check: “Here is an objection and my reply. Does the reply answer it, concede it or change the subject? Point to the relevant sentence.”
- Strongest alternative: “Suggest a plausible alternative to my proposal. Ask me to explain why my approach is better under the same constraints.”
- Cross-question practice: “Ask three short questions about my assumptions, one at a time. Accept ‘I do not know’ and flag what needs research.”
Prompts for listening and revision
- Switch sides: “Help me write the strongest reasonable opposing case. Preserve the original opponent’s scope and ask before assigning them a belief.”
- Common-ground worksheet: “From these two statements, distinguish shared goals, shared factual assumptions and unresolved disagreements. Quote only words actually provided.”
- Evidence audit: “List factual claims in this draft. Do not verify them from memory. Label each as supported by the supplied source, unsupported or a proposal.”
- Revision drill: “Pick one weak paragraph and offer two clearer versions. Explain what each changes without adding new evidence or declaring a winner.”
Example: turn an evasive reply into a relevant reply
A useful assistant exercise is checking alignment between objection and answer. The following exchange is hypothetical.
How to assess the assistant’s output
Keep a copy of your original draft and compare what changed. Do not accept a rewrite simply because it sounds more forceful.
Swipe across to compare all columns →
| Check | Useful result | Warning sign |
|---|---|---|
| Faithfulness | Keeps your scope and uncertainty | Turns “might” into “will” |
| Evidence | Separates supplied facts from guesses | Adds a specific uncited statistic |
| Opposing view | Represents a recognizable concern | Invents an extreme position |
| Actionability | Names a sentence to revise | Gives only a vague score |
Move the revised argument into a real exchange
Use the prompt output to prepare one improved opening and one honest question. Then discuss the issue with a person. Tell them that you rehearsed; do not pretend the assistant’s simulated agreement is evidence of public support. A shared-device Debagle session can give that human exchange a turn structure.
Questions people ask
Do these prompts require browsing?
No. They are writing and practice exercises. If you ask an assistant to research facts, separately confirm that sources are accessible and actually support the claim.
Should I ask the AI to win at any cost?
That encourages rhetorical tricks rather than useful practice. Ask it to challenge reasoning accurately and identify uncertainty.
Can I paste a private transcript?
Get permission from participants and check the service’s current data terms. For practice, a short anonymized hypothetical is often enough.
