research

---

name: research description: LLM Council - 3-stage multi-model research pipeline (collect, peer review, synthesize). Queries OpenAI, Perplexity, Claude, and Gemini in parallel, has them anonymously critique each other, then synthesizes a final answer. Keywords: research, code review, architecture, multi-model consensus, LLM council, peer review, technology comparison.

Research Skill

LLM Council: 3-stage multi-model research using OpenAI, Perplexity, Claude, and Gemini.

How It Works

1. Collect - Query all models in parallel
2. Peer Review - Each model anonymously critiques and ranks the others
3. Synthesize - Chairman produces a final answer from all responses and reviews

9 API calls per query with all 4 models (4 collect + 4 review + 1 synthesis). Gracefully degrades if models are unavailable.

The chairman is selected based on task type: Perplexity leads research queries, Gemini leads code reviews, bug analysis, and comparisons.

Usage Modes

Monolithic (full council pipeline)

Runs all 3 stages with every available provider:

python -m research research "Python async best practices"
python -m research review src/server.py "security focus"
python -m research compare LanceDB Pinecone Weaviate
python -m research bug src/client.py "Connection timeout"
python -m research design src/arch.py "scalability requirements"
python -m research validate "Use DuckDB" "analytics pipeline" PostgreSQL ClickHouse
python -m research comprehensive src/auth.py "authentication best practices"

Per-agent (single provider query)

Query a single provider directly. Used by model-agent definitions in multi-agent workflows:

python -m research query openai "Python async best practices"
python -m research query gemini "Compare DuckDB and PostgreSQL" --task-type comparison
python -m research query perplexity "Latest WASM runtime benchmarks" --context "Performance research"
python -m research query anthropic "Review this approach" --task-type code-review

Peer review (on collected responses)

Run peer review on a JSON file of previously collected responses:

python -m research review-responses "Python async best practices" responses.json

The responses file should be a JSON array of objects with provider, model, and content fields.

Programmatic

import asyncio
from research import run_council, query_single

# Full council pipeline
result = asyncio.run(run_council("Python async best practices", "Research query"))
print(result.synthesis)

# Single provider query
response = asyncio.run(query_single("gemini", "Explain WASM", "Research query"))
print(response.content)

Setup & Verification

See SETUP.md for detailed provider configuration, agent symlink verification, and troubleshooting. When users ask about configuring this skill or debugging provider issues, refer to that document.

Environment

# Required (at least one)
OPENAI_API_KEY=sk-proj-...
PERPLEXITY_API_KEY=pplx-...
ANTHROPIC_API_KEY=sk-ant-...

# Gemini via Vertex AI
GOOGLE_CLOUD_PROJECT=your-gcp-project-id
GOOGLE_CLOUD_LOCATION=global          # CRITICAL: use "global" for Gemini 3.x
GOOGLE_GENAI_USE_VERTEXAI=TRUE
GOOGLE_APPLICATION_CREDENTIALS=/path/to/service-account.json

# Override models (optional — defaults are set in council.py)
OPENAI_MODEL=<model-id>
PERPLEXITY_MODEL=<model-id>
ANTHROPIC_MODEL=<model-id>
GEMINI_MODEL=<model-id>

Multi-Agent Research Team

This skill supports two orchestration modes:

CLI mode (single process)

Run python -m research research "topic" to execute all 3 stages in one process. Simple but inflexible — always uses the same models regardless of task type.

Team mode (recommended)

The research-lead agent orchestrates a team of specialized agents. All agents inherit their model from ~/.claude/settings.json.

Agent	Role
research-lead	Orchestrator: decomposes queries, selects which model-agents to spawn, synthesizes
openai-researcher	Queries OpenAI via query openai — chain-of-thought scaffolding
perplexity-researcher	Queries Perplexity via query perplexity — citation-backed answers
claude-researcher	Answers directly (no API key needed) — nuanced synthesis with codebase access
gemini-researcher	Queries Gemini via query gemini — multimodal reasoning
peer-reviewer	Evaluates anonymized responses via review-responses — avoids self-review bias

The research-lead selects agents based on task type:

Task	Agents Selected
research	All 4 researchers
review	OpenAI, Claude, Gemini
compare	Perplexity, Claude, Gemini
bug	OpenAI, Claude
validate	All 4 researchers

Dependencies

httpx and google-genai. Install: pip install httpx google-genai