The Council

Developers often struggle with AI "yes-man" bias, where models reflexively agree with a user's stated preference or plan. This skill forces objective scrutiny by convening a multi-model council that operates through a hardened workflow to ensure honesty.

What it does

• Multi-model voting triggers independent ballots from Claude Opus, Sonnet, and Fable to eliminate single-model blind spots.
• Blind framing rewrites user requests into neutral, symmetric questions so voters cannot detect the user's desired outcome.
• Premise verification uses automated tools to audit load-bearing claims before a verdict is reached.
• Red teaming automatically triggers a dissent round if the council unanimously agrees with the user on high-stakes decisions.
• Outside seats allows optional integration of GPT or Gemini models to provide vendor-diverse perspectives.

How it works

1. Intake involves parsing your question, identifying your stated stance, and gathering necessary repository context.
2. Neutralization occurs via a framer agent that shuffles options and removes evaluative language to prevent anchoring.
3. Balloting runs independent seats with verification tool access to check facts against the codebase or web.
4. Adjudication by a non-voting chair rules on evidence and compiles a minority report for any dissenting views.

Frameworks & tools

This skill utilizes the Claude Code Workflow tool, jq for result parsing, and local shell scripts for optional third-party model integration.

Why this beats prompting it yourself

Standard prompting fails to hide user intent from the model, leading to sycophancy. This skill uses a programmatic workflow to physically separate the "framer" from the "voters," making it impossible for the judging models to simply tell you what you want to hear.

Use cases

• Architecture reviews to determine if a proposed migration or refactor is genuinely the best path.
• Sanity checks for complex debugging theories where you might be chasing a red herring.
• Strategic decisions where you need a "devil's advocate" to poke holes in a business or technical plan.
• Security assessments to red-team a proposed auth flow or permission model.

Known limitations

Requires Claude Code or a compatible Workflow runner. Outside seats (GPT/Gemini) require local CLI configuration and existing API billing. Deep mode may consume significant token quotas.