Create Perspective Diversity with Specialized Agent Prompts

The council assigns different prompts, perspectives, and biases to its agents so that the decision is evaluated from multiple angles. One configured role set includes a contrarian, a first-principles agent, an expansionist, an outsider, an executor, and a final chairman who makes the final call.

Additional run labels include pragmatist, red teamer, generalist, and reviewer. Across these variations, the mechanism is the same: agents do not all receive an identical role prompt.

Each agent is spawned as a separate sub-agent. Its prompt defines the perspective it should bring to the decision. Because the prompts differ, the agents are expected to produce distinct analyses even when they run on the same model.

The council then uses a final decision role, described as a chairman, along with reviewer stages. This separates contribution from synthesis. The agents generate perspectives; the final stage weighs them and produces the answer.

The roles are not presented as independent external experts. They are configured AI agents. Their diversity comes from deliberate prompting and, in the advanced mode, from using different underlying models.

Role specialization prevents the council from becoming several copies of the same generic assistant prompt. A contrarian configuration, a first-principles configuration, an outsider configuration, and an execution-oriented configuration are intended to approach the same question differently.

This is how a single-model council can still implement stochastic debate. The underlying model is shared, but the system varies the instructions that govern each agent's contribution.

After launching a council run, inspect the sub-agent entries rather than looking only at the final response. Confirm that:

• Multiple agents were spawned.
• Each agent has its own prompt.
• The roles are meaningfully separated rather than identical.
• Reviewer or final-decision agents appear after the initial analyses.
• The council completes its synthesis before presenting the final recommendation.

This inspection is especially relevant during setup. It verifies that the skill is invoking the intended multi-agent workflow instead of falling back to a normal single-response chat.

Role diversity is not the same as model diversity. When every agent uses the same model, all agents may still share that model's characteristic response patterns and vulnerabilities. Different prompts can broaden the analysis, but they do not remove the fact that the same underlying model is generating every perspective.

The role names identify separate prompts and biases, but the full prompt text is not reproduced here. The setup therefore assumes that the supplied skill files already contain the role definitions.

A council needs distinct agent prompts and a separate synthesis role. Prompt diversity creates multiple perspectives within one model; model diversity, introduced later, adds another layer by changing the underlying model itself.