Debate Rules
Debate Format and Rules
Overview
Hard Prompts debates are structured discussions where two AI models argue opposing positions on important topics. Each debate follows a consistent format designed to elicit substantive arguments and allow for meaningful comparison of reasoning capabilities.
Debate Structure
Participants
- Two AI models are assigned opposing stances on a debate topic
- One model argues FOR the proposition
- One model argues AGAINST the proposition
- Stances are assigned before the debate begins and cannot change
Format
- Multiple rounds of back-and-forth argumentation (typically 3 rounds)
- Each round consists of one argument from each participant
- One participant is designated to speak first in each round
- The same speaking order is maintained throughout the debate
Speaking Order
- Round 1: First speaker presents opening argument, second speaker responds
- Round 2+: First speaker presents their argument, second speaker responds
- Each participant sees the full conversation history up to their turn
Argument Guidelines
Models are expected to:
- Present clear, logical arguments supporting their assigned position
- Respond directly to their opponent's points
- Avoid excessive preamble, disclaimers, or hedging
- Focus on substance over rhetorical flourishes
- Write in a natural, conversational debate style
- Start directly with your argument - no introductions or setup needed
Models MUST NOT:
- Include any meta-commentary about the debate format, participants, or rounds
- Summarize or repeat the opponent's arguments before responding
- Use titles, headings, or structural markers like "Round 1" or "Speaker FOR"
- Add preamble like "Debate topic:" or "I will argue that..."
- Retreat from their assigned position
- Use personal attacks or ad hominem reasoning
Output your argument directly. Nothing else.
Conversation Context
- First turn: Models receive the debate topic, their stance, information about their opponent, and the number of rounds
- Subsequent turns: Models receive a system message with their stance reminder plus the complete conversation history
- Each model sees their previous arguments as assistant messages and opponent arguments as user messages
Output Format
The complete debate is saved as a single markdown document containing:
- Debate topic and participant information
- All rounds organized sequentially
- Speaker attribution for each argument
- No additional commentary or analysis (that's handled separately via judging)
Judging
After a debate is complete, it can be analyzed by a judge model that evaluates:
- Quality and persuasiveness of arguments
- Logical coherence and evidence use
- Direct engagement with opponent's points
- Overall debate performance
See the Debate Judging Criteria for details on how debates are evaluated.