AgentSkillsCN

llm-research

通过并行运行多个 LLM 模型(Codex、Claude、Gemini),对同一研究问题展开独立分析,并收集各自的见解。适用于游戏设计评审、架构分析、代码审查,或任何能够从多元视角中获得价值的研究任务。

SKILL.md
--- frontmatter
name: llm-research
description: >
  Run a research question through multiple LLM models (Codex, Claude, Gemini) in parallel
  and collect their independent analyses. Use for game design reviews, architecture analysis,
  code reviews, or any research task where diverse perspectives add value.

LLM Research Skill

Quick Start

Run research from 3 models in parallel:

bash
python3 scripts/llm_research.py "Your research question here"

Or with a spec file:

bash
python3 scripts/llm_research.py --spec /path/to/research-spec.json

Workflow

  1. Prepare prompt: Write a clear research question with expected output format
  2. Run script: Launch all 3 models in parallel
  3. Wait: Codex (xhigh reasoning) takes 2-5 minutes, others are faster
  4. Collect: Results saved to ./llm-research/runs/<timestamp>/
  5. Synthesize: Optionally combine insights into a summary

Output Structure

code
llm-research/runs/YYYYMMDD-topic/
├── research-codex.md      # Codex (gpt-5.2-codex, xhigh reasoning)
├── research-claude.md     # Claude Opus
├── research-gemini.md     # Gemini Pro Preview
└── synthesis.md           # Combined insights (optional)

Spec File Format

json
{
  "topic": "Game Design Review",
  "prompt": "Analyze the game mechanics and provide...",
  "context_files": ["docs/design/GAME_DESIGN.md"],
  "output_format": "## 1. Section\n## 2. Section\n...",
  "synthesize": true
}

CLI Options

  • --spec FILE - Use spec file instead of inline prompt
  • --output-dir DIR - Custom output directory
  • --no-codex - Skip Codex (faster, 2 models only)
  • --no-synthesis - Don't generate combined summary
  • --timeout MINUTES - Max wait time (default: 10)

Model Configuration

Uses ~/.config/llm-council/agents.json for model settings:

json
{
  "planners": [
    { "name": "codex-1", "kind": "codex", "model": "gpt-5.2-codex", "reasoning_effort": "xhigh" },
    { "name": "claude-2", "kind": "claude", "model": "opus" },
    { "name": "gemini-3", "kind": "gemini", "model": "gemini-3-pro-preview" }
  ]
}

Example: Game Design Review

bash
cat > /tmp/review-spec.json << 'EOF'
{
  "topic": "Investomania Cards Review",
  "prompt": "You are an expert card game designer. Read docs/design/GAME_DESIGN.md and provide a structured review in Russian...",
  "context_files": ["docs/design/GAME_DESIGN.md"],
  "output_format": "## 1. Общая оценка\n## 2. Анализ механик\n## 3. Проблемы\n## 4. Улучшения\n## 5. Сравнение\n## 6. Плейтесты",
  "synthesize": true
}
EOF

python3 scripts/llm_research.py --spec /tmp/review-spec.json

Key Differences from LLM Council

AspectLLM CouncilLLM Research
OutputPlans (how to do)Answers (done)
JudgeYes (scores plans)Optional synthesis
Use caseImplementation planningAnalysis, reviews
Time5-30 min2-10 min

Best Practices

  1. Be specific: Give concrete output format, not vague instructions
  2. Include context: Reference files the models should read
  3. Set expectations: "Give numbers", "Be critical", "Minimum 5 points"
  4. Russian output: Explicitly state language in prompt

Constraints

  • Models run independently (no shared context)
  • Codex with xhigh reasoning is slowest but most thorough
  • Total token budget ~50k per model
  • Don't expect identical outputs — diversity is the point