council-refine

/council-refine — Multi-Model Plan Refinement with Opus 4.6 Synthesis

Submit a plan to 3 frontier AI models (GPT-5.2, Claude Sonnet 4.5, Gemini 3 Pro) via Perplexity API for iterative refinement, with Opus 4.6 extended thinking synthesis. Loop until convergence (score >= 8, no critical issues, score gain < 1) or max 3 iterations.

Prerequisites: PERPLEXITY_API_KEY and ANTHROPIC_API_KEY environment variables set. Python packages installed (pip install perplexityai anthropic).

Input

$ARGUMENTS = The plan text to refine. Can be multi-line.

Workflow

Step 0: Initialize

• Store the plan text from $ARGUMENTS
• Set iteration = 1, maxIterations = 3, previousScore = 0

Step 1: Build refinement query

Compose the council query for this iteration:

You are a panel of expert reviewers evaluating a technical implementation plan. Analyze the following plan thoroughly and provide:

1. OVERALL SCORE (1-10): How production-ready is this plan?
2. STRENGTHS: What's well-designed?
3. WEAKNESSES: What needs improvement?
4. CRITICAL ISSUES: Blockers that must be fixed before implementation
5. SPECIFIC IMPROVEMENTS: Concrete, actionable changes with code/architecture suggestions
6. REVISED SECTIONS: Rewritten sections incorporating your feedback

PLAN TO REVIEW (Iteration {iteration}/{maxIterations}):
---
{planText}
---

{If iteration > 1: "PREVIOUS FEEDBACK ADDRESSED: {summary of changes made}"}

IMPORTANT: Include a numeric score (1-10) prominently in your response.

Step 2: Run council query

Call council_query MCP tool with:

• query: The prompt from Step 1
• mode: "browser" (Playwright browser automation — no API keys needed, uses Perplexity login)
• includeContext: true

This runs the full pipeline: 3 parallel Perplexity queries + Opus 4.6 synthesis with extended thinking (~20s total).

Step 3: Read and evaluate

The council_query response contains the Opus synthesis. Extract:

• Score: From recommended_actions or narrative — look for the numeric score
• Critical issues: From risks and disagreements
• Improvements: From recommended_actions and unique_insights