Skip to content

USAGE_GUIDE.md

Latest commit

 

History

History
269 lines (203 loc) · 7.85 KB

USAGE_GUIDE.md

File metadata and controls

269 lines (203 loc) · 7.85 KB

🎭 LLM Critique Tool - Usage Guide

Quick Start

# Expert creator + Expert critics (most powerful)
python -m llm_critique.main "Design a mobile app" --creator-persona steve_jobs --personas ray_dalio

# Simple vanilla setup
python -m llm_critique.main "Review my code" --creator-persona gpt-4o --critique-models claude-4-sonnet,grok-beta

Expert Personas vs 🤖 Vanilla Models

Expert Personas

  • Rich personality: Authentic voice, expertise, thinking patterns
  • Examples: steve_jobs, ray_dalio, elon_musk, warren_buffett
  • Features: Domain expertise, decision frameworks, characteristic language

Vanilla Models

  • Standard AI: Clean, professional responses without personality
  • Examples: gpt-4o, claude-4-sonnet, gemini-2.0-flash, grok-beta, o3-mini
  • Features: Cost-efficient, reliable, general-purpose

📋 Complete Usage Patterns

1️⃣ Expert Creator + Expert Critics

Best for: Rich, authentic content with specialized domain critique

python -m llm_critique.main "Your prompt" \
  --creator-persona elon_musk \
  --personas steve_jobs,ray_dalio

What happens: Elon Musk creates content with his characteristic first-principles thinking, while Steve Jobs focuses on user experience and Ray Dalio provides systematic analysis.

2️⃣ Expert Creator + Expert Critics (Global Model Override)

Best for: Expert personalities but forcing a specific model for cost/speed control

python -m llm_critique.main "Your prompt" \
  --creator-persona elon_musk \
  --personas steve_jobs,ray_dalio \
  --personas-model o3-mini

What happens: All expert personas use o3-mini instead of their preferred models, maintaining personality but with consistent model behavior.

3️⃣ Expert Creator + Vanilla Critics

Best for: Authentic creator voice with cost-efficient, general-purpose critique

python -m llm_critique.main "Your prompt" \
  --creator-persona steve_jobs \
  --critique-models gpt-4o,claude-4-sonnet

What happens: Steve Jobs creates with his design philosophy, while vanilla models provide professional, unbiased critique.

4️⃣ Vanilla Creator + Expert Critics

Best for: Professional content creation with specialized expert review

python -m llm_critique.main "Your prompt" \
  --creator-persona gpt-4o \
  --personas ray_dalio,warren_buffett \
  --personas-model claude-4-sonnet

What happens: Clean, professional content creation followed by expert business analysis and investment perspective.

5️⃣ Vanilla Creator + Vanilla Critics

Best for: Cost-efficient, fast, professional workflow

python -m llm_critique.main "Your prompt" \
  --creator-persona gpt-4o \
  --critique-models claude-4-sonnet,gemini-2.0-flash

What happens: Standard AI content creation with multi-model professional critique.

6️⃣ All Expert Personas (Global Model)

Best for: Maximum expertise diversity with model control

python -m llm_critique.main "Your prompt" \
  --creator-persona elon_musk \
  --personas all \
  --personas-model o3-mini

What happens: Uses ALL available expert personas as critics, all running on the specified model.

🔧 Key Rules & Constraints

Mutually Exclusive Options

# ❌ ERROR - Cannot mix these
--personas steve_jobs --critique-models gpt-4o

# ✅ CORRECT - Choose one approach
--personas steve_jobs                    # Expert personas
--critique-models gpt-4o                 # Vanilla models

Global Model Override Rules

# ✅ WORKS - Overrides all persona models
--personas steve_jobs,ray_dalio --personas-model o3-mini

# ❌ ERROR - personas-model only works with --personas
--critique-models gpt-4o --personas-model o3-mini

# ❌ ERROR - "all" requires global model
--personas all

Creator Options

# Expert creator (rich personality)
--creator-persona steve_jobs

# Vanilla creator (standard AI)
--creator-persona gpt-4o

# Auto-select (uses first critic's model)
# No --creator-persona flag

📖 Discovery Commands

# See all expert personas and vanilla models
python -m llm_critique.main --list-personas

# See all supported AI models and API requirements
python -m llm_critique.main --list-models

# Get detailed info about a specific persona
python -m llm_critique.main --persona-info steve_jobs

# Estimate cost before running
python -m llm_critique.main "Your prompt" --creator-persona steve_jobs --personas ray_dalio --est-cost

💡 Common Scenarios

Business Strategy Analysis

python -m llm_critique.main "Launch strategy for AI product" \
  --creator-persona elon_musk \
  --personas ray_dalio,warren_buffett \
  --personas-model o3-mini

Product Design Review

python -m llm_critique.main "Mobile app wireframes feedback" \
  --creator-persona steve_jobs \
  --personas simon_sinek

Technical Architecture Review

python -m llm_critique.main "Microservices architecture proposal" \
  --creator-persona mark_russinovich \
  --critique-models claude-4-sonnet,gemini-2.0-flash

Cost-Efficient General Review

python -m llm_critique.main "Review my blog post" \
  --creator-persona gpt-4o-mini \
  --critique-models claude-3-haiku,gemini-2.0-flash

Maximum Expert Insight

python -m llm_critique.main "Evaluate this startup idea" \
  --creator-persona mark_andreessen \
  --personas all \
  --personas-model claude-4-sonnet

⚠️ Common Mistakes

❌ Mixing Persona Types

# ERROR: Cannot mix --personas and --critique-models
--personas steve_jobs --critique-models gpt-4o

❌ Wrong Global Model Usage

# ERROR: --personas-model only works with --personas
--critique-models gpt-4o --personas-model claude-4-sonnet

❌ Missing Required Model

# ERROR: --personas all requires --personas-model
--personas all

🎯 Tips for Best Results

  1. Match creator to content type: Use Steve Jobs for product design, Elon Musk for technical innovation, Ray Dalio for strategy
  2. Consider cost vs quality: Expert personas cost more tokens but provide richer insights
  3. Use global model override: --personas-model for consistent behavior across experts
  4. Mix approaches: Expert creator + vanilla critics for balanced cost/quality
  5. Start with cost estimation: Use --est-cost to understand token usage before running

🔧 Advanced Features

Multiple Iterations

--iterations 3  # Run 3 creator-critic cycles for refinement

Save Conversations

--listen my_session.txt  # Save rich conversation log

JSON Output

--format json  # Machine-readable output

Cost Estimation

--est-cost  # Estimate tokens and cost without running

🚀 X AI Grok Models

X AI's Grok models offer real-time access to X (Twitter) data and advanced reasoning capabilities:

Available Models

  • grok-beta: Current production model with real-time X data integration
  • grok-3: Latest flagship model with 1M context window (when available)
  • grok-3-mini: Faster, cost-efficient version
  • grok-3-reasoning: Advanced reasoning with "Think" mode
  • grok-2: Previous generation model

Setup

export XAI_API_KEY="your-xai-api-key"

Usage Examples

# Latest Grok with real-time data
python -m llm_critique.main "Analyze recent tech trends" --creator-persona grok-beta --critique-models grok-3-reasoning

# Cost-efficient setup
python -m llm_critique.main "Quick code review" --creator-persona grok-3-mini --critique-models grok-beta

# Reasoning-focused analysis
python -m llm_critique.main "Complex math problem" --creator-persona grok-3-reasoning --critique-models steve_jobs --personas-model grok-3

🎯 Common Usage Patterns

Need help? Use python -m llm_critique.main --help for quick reference or --list-personas to explore available experts!