Objective: test a plan, a design or a project idea through multi-agent debate, starting with a silent "grill me" framing, then bringing the models to converge on a useful and critical synthesis.
Principles
- Each LLM receives the same initial brief.
- Each LLM first works independently.
- The responses of the other LLMs are introduced only after the first position.
- Each round follows a stable format.
- The debate prioritizes clarity, useful criticism and the detection of blind spots.
- Do not invent facts absent from the brief.
- Do not make global recommendations before the synthesis phase.
- The final judge arbitrates divergences and produces a consolidated decision.
Round 0 — Silent framing
Goal
Mentally activate the "grill me" logic before any visible debate.
Instructions
- Read the brief.
- Mentally execute the "grill me" skill.
- Identify the decision branches.
- Identify the dependencies between branches.
- Identify ambiguities, implicit assumptions and risk areas.
- Prepare an initial position.
- Do not yet produce a full critique.
Expected output
## Round 0 — Cadrage silencieux
### Branches de décision
- ...
### Dépendances
- ...
### Ambiguïtés
- ...
### Questions prioritaires
- ...
### Préparation pour Round 1
- ...
Round 1 — Initial position
Goal
Give an autonomous analysis without taking the other responses into account.
Instructions
- Produce an independent initial analysis.
- Do not take the other responses into account.
- Be critical but useful.
- Distinguish facts, assumptions, risks and recommendations.
- Prioritize blockers before minor improvements.
Expected output
## Round 1 — Position initiale
### Résumé du problème
- ...
### Hypothèses implicites
- ...
### Points forts
- ...
### Points faibles
- ...
### Risques critiques
- ...
### Questions de clarification
- ...
### Recommandations provisoires
- ...
### Verdict provisoire
- ...
Round 2 — Cross-reading
Goal
React to the other LLMs' responses.
Instructions
- Read the other LLMs' positions.
- Say what is kept, corrected, rejected.
- Explain the important divergences.
- Revise the analysis if necessary.
- Keep the still-open risks visible.
- Do not settle for saying "I agree".
Expected output
## Round 2 — Lecture croisée
### Ce que je conserve
- ...
### Ce que je corrige
- ...
### Ce que je rejette
- ...
### Divergences importantes
- ...
### Révision de mon analyse
- ...
### Risques toujours ouverts
- ...
Round 3 — Final revision
Goal
Produce a consolidated version after the debate.
Instructions
- Integrate the relevant criticisms.
- Reduce blind spots.
- Update the confidence level.
- Keep the remaining uncertainties visible.
- Produce actionable final recommendations.
Expected output
## Round 3 — Révision finale
### Position consolidée
- ...
### Ce qui a changé
- ...
### Risques restants
- ...
### Points encore incertains
- ...
### Recommandations finales
- ...
### Verdict final provisoire
- ...
Final judge
Goal
Synthesize the debate and arbitrate.
Instructions
- Compare the arguments.
- Identify convergences.
- Flag structural divergences.
- Identify the strongest points.
- Decide if possible.
- Flag what remains uncertain.
- Do not count votes only.
Expected output
## Juge final
### Synthèse des convergences
- ...
### Synthèse des divergences
- ...
### Arguments les plus solides
- ...
### Points de vigilance majeurs
- ...
### Décision finale
- ...
### Niveau de confiance
- ...
### Ce qu'il faut faire ensuite
- ...
Degraded mode
If the debate is interrupted before the end:
- produce a partial report;
- list the completed rounds;
- list the missing rounds;
- list the recommendations already issued;
- explicitly state that the final synthesis did not take place.
Partial report format
## Rapport partiel
### Rounds complétés
- ...
### Rounds restants
- ...
### Recommandations déjà émises
- ...
### Limites de validité
- ...
Warning
This protocol tests the internal consistency of a plan or design, not its external validity or its real feasibility.