PM Prompt Toolkit

Production-grade prompt patterns and multi-cloud AI orchestration for product teams building with Claude, GPT, and Gemini. Now with AWS Bedrock and Google Vertex AI support. Proven at scale: 5K+ signals/week, 95% accuracy, $0.001/signal cost.

📊 Project Status

Status: Production-ready | Last Updated: 2025-10-27 | Maintenance: Active development

---

🎯 Why This Exists

Problem: Most prompt libraries showcase toy examples. Production systems need battle-tested patterns with real metrics.

Solution: Enterprise-grade toolkit combining 200+ production prompts, multi-model orchestration, and cost optimization strategies.

Key Benefits:

• Proven ROI: 99.7% cost reduction through intelligent model cascading (Haiku → Sonnet → Opus)
• Production Metrics: 95% accuracy on 5K+ weekly signals, validated on $100M+ ARR systems
• Multi-Model Expertise: Optimized patterns for Claude 4.x, GPT-4o, Gemini 2.5 families
• Zero to Production: Complete examples with evaluation methodology, not just prompts
• Developer Tools: Python package with YAML-based model registry, pricing service, capability validation

---

🚀 Quick Start

Get operational in under 5 minutes:

# Clone and install
git clone https://github.com/awoods187/PM-Prompt-Patterns.git
cd PM-Prompt-Patterns
pip install -e .

# Verify installation
python -c "from ai_models import get_model; print(get_model('claude-sonnet-4-5').name)"

Next Steps by Role:

• New to prompts? → Design Principles
• Optimizing costs? → Model Selection Guide
• Building production systems? → Epic Categorization Example
• Understanding the codebase? → Project Structure

---

📋 Prerequisites

Tool	Version	Verification	Purpose
Python	3.9+	python --version	Core runtime
pip	Latest	pip --version	Package management
API Keys	-	Set in .env	Claude/GPT/Gemini access

API Key Setup (Optional)

Create .env file:

ANTHROPIC_API_KEY=your_key_here
OPENAI_API_KEY=your_key_here
GOOGLE_API_KEY=your_key_here

Keys only required for live API testing. Browse prompts without keys.

Multi-Provider Support (New in v0.2.0)

The toolkit supports 6 AI providers with intelligent routing and cost optimization:

Supported Providers:

• Anthropic Claude (direct API) - Haiku, Sonnet, Opus
• AWS Bedrock - Claude models via AWS infrastructure
• Google Vertex AI - Claude models via Google Cloud
• OpenAI - GPT-5 (all variants), GPT-4.1, o3, o4-mini, GPT-4o
• Google Gemini - Gemini 2.5 Pro, Flash, Flash Lite
• Mock Provider - Zero-cost testing provider

Installation:

# Base installation (Anthropic Claude + Mock)
pip install -e .

# With AWS Bedrock support
pip install -e ".[bedrock]"

# With Google Vertex AI support
pip install -e ".[vertex]"

# With OpenAI support
pip install -e ".[openai]"

# With Google Gemini support
pip install -e ".[gemini]"

# With all providers
pip install -e ".[all]"

Provider Configuration (.env file):

# Anthropic Claude (direct API)
ANTHROPIC_API_KEY=your_key_here

# AWS Bedrock
ENABLE_BEDROCK=true
AWS_ACCESS_KEY_ID=your_key_here
AWS_SECRET_ACCESS_KEY=your_secret_here
AWS_REGION=us-east-1

# Google Vertex AI
ENABLE_VERTEX=true
GCP_PROJECT_ID=your-project-id
GCP_REGION=us-central1
GCP_CREDENTIALS_PATH=/path/to/credentials.json  # Optional

# OpenAI
ENABLE_OPENAI=true
OPENAI_API_KEY=sk-your_key_here
OPENAI_ORG_ID=org-your_org_here  # Optional

# Google Gemini
GOOGLE_API_KEY=your_key_here

Usage Examples:

from pm_prompt_toolkit.providers import get_provider

# Explicit provider selection with prefix
claude = get_provider("anthropic:claude-sonnet-4-5")
bedrock = get_provider("bedrock:claude-sonnet-4-5")
vertex = get_provider("vertex:claude-sonnet-4-5")
openai = get_provider("openai:gpt-5")  # GPT-5 flagship with auto-routing
gemini = get_provider("gemini:gemini-2-5-pro")

# Automatic routing (Claude models prefer Bedrock if enabled)
provider = get_provider("claude-sonnet-4-5")  # Uses Bedrock if enabled
result = provider.classify("We need SSO integration")

# Direct model routing
openai_provider = get_provider("gpt-5")  # Routes to OpenAI if enabled (auto-routing Instant/Thinking)
gemini_provider = get_provider("gemini-2-5-flash")  # Routes to Gemini

# Mock provider for testing (zero cost)
mock = get_provider("mock:claude-sonnet")

Provider Benefits:

• Bedrock: Enterprise AWS infrastructure, AWS-native billing, regional data residency
• Vertex AI: Google Cloud integration, GCP-native billing, unified GCP experience
• OpenAI: Advanced function calling, structured outputs, multimodal
• Gemini: 2M token context (Pro), ultra-low cost (Flash Lite), context caching
• Fallback: Automatically falls back to direct Anthropic API if cloud providers unavailable

---

📖 Usage

Basic Model Management

from ai_models import get_model, has_vision, has_prompt_caching

# Get model specifications
model = get_model("gpt-5")  # GPT-5 flagship
print(f"Context: {model.metadata.context_window_input:,} tokens")
print(f"Cost: ${model.pricing.input_per_1m}/M input tokens")

# Calculate costs with caching
cost = model.calculate_cost(
    input_tokens=10_000,
    output_tokens=2_000,
    cached_input_tokens=5_000  # 90% discount on cached tokens
)

# Validate capabilities before API calls
if has_vision("gpt-5"):  # GPT-5 has vision capabilities
    process_image()

Finding Budget-Friendly Models

from ai_models import ModelRegistry

budget_models = ModelRegistry.filter_by_cost_tier("budget")
# Returns: Haiku 4.5, GPT-5 mini, GPT-4.1 mini, Gemini 2.5 Flash

Advanced Examples: API Documentation | Production Architecture

Provider-Optimized Prompts ✨ ALL PROMPTS MIGRATED

All 13 production prompts now have provider-specific optimizations for Claude, OpenAI, and Gemini:

from ai_models import get_prompt, get_model, list_prompts

# See all available prompts
prompts = list_prompts()
# Returns: ['analytics/signal-classification', 'developing-internal-tools/claude-md-generator', ...]

# Automatically select best prompt variant for your model
model = get_model("gpt-5")  # GPT-5 default model
prompt = get_prompt("analytics/signal-classification", model=model.id)

# Or explicitly choose a provider optimization
claude_prompt = get_prompt("developing-internal-tools/code-review-refactoring", provider="claude")
openai_prompt = get_prompt("product-strategy/meta-prompt-designer", provider="openai")
gemini_prompt = get_prompt("stakeholder-communication/executive-deck-review", provider="gemini")

Provider-Specific Features:

Provider	Key Optimizations	Best For	Cost/1K Operations
Claude	XML tags, chain-of-thought, caching	Complex reasoning, accuracy	$1-15
OpenAI	Auto-routing modes, reasoning (o3/o4), function calling, JSON mode	Coding (4.1), reasoning (o3), general tasks (GPT-5)	$0.15-40
Gemini	2M context, caching, batch processing	High volume, cost optimization	$0.038-5

All 13 Prompts Available:

• Analytics: signal-classification
• Development: claude-md-generator, code-review-refactoring, enterprise-readme-generator, github-actions-python-cicd, llm-orchestration-system, prompt-extraction-cataloging, pytest-cicd-optimization, python-80-percent-test-coverage
• Strategy: meta-prompt-designer, opus-code-execution-pattern
• Communication: executive-deck-review, remove-ai-writing-patterns

Total: 13 prompts × 4 variants (base + Claude + OpenAI + Gemini) = 52 optimized prompt variants

→ Learn more about provider-optimized prompts | → Migration tool

---

🤖 Model Selection Guide

Choose the right model for your workload:

Model	Input/Output (per 1M)	Context	Best For	Our Usage
Claude Haiku 4.5	$1/$5	200K	High-volume classification	70%
Claude Sonnet 4.5	$3/$15	200K	Production workhorse	25%
Claude Opus 4.1	$15/$75	200K	High-stakes decisions	5%
GPT-5	$3/$12	256K	Flagship auto-routing (Instant/Thinking modes)	Default
GPT-5 mini	$0.20/$0.80	128K	Fast, efficient GPT-5 variant	Budget
GPT-4.1	$2.5/$10	200K	Specialized for coding, precise instructions	Coding
GPT-4.1 mini	$0.15/$0.60	128K	Fast coding model, replaced 4o-mini	Budget coding
o3	$10/$40	128K	Advanced reasoning, full tool access	Complex reasoning
o4-mini	$1/$4	128K	Fast, cost-efficient reasoning	Math/coding
GPT-4o	$2.5/$10	128K	Multimodal specialist (voice, vision)	Multimodal
Gemini 2.5 Pro	$1.25/$5	2M	Massive context analysis	Large docs
Gemini 2.5 Flash	$0.075/$0.30	1M	Speed-critical apps	Real-time

Pricing verified October 2025. See MODEL_OPTIMIZATION_GUIDE.md for detailed comparison.

Cost Optimization Pattern

Don't use one model for everything. Intelligent routing saves 99.7%:

Keyword Filter (free) → 70% resolved
    ↓
Haiku ($0.0003/signal) → 25% resolved
    ↓
Sonnet ($0.002/signal) → 4.5% resolved
    ↓
Opus ($0.015/signal) → 0.5% resolved

Average: $0.001/signal (vs $0.015 naive approach)

→ Implementation Guide

---

📁 Repository Structure

PM-Prompt-Patterns/
├── prompts/              # Production-ready prompts by category
│   ├── analytics/        # Monitoring, reporting, investigation (MECE)
│   ├── product-strategy/ # Roadmapping, prioritization
│   └── technical-docs/   # API docs, CLAUDE.md generation
├── ai_models/            # Python model management system
│   ├── registry.py       # YAML-based model registry
│   ├── pricing.py        # Cost calculation with caching
│   └── definitions/      # Model specs (Anthropic, OpenAI, Google)
├── examples/             # Complete production systems
│   └── epic-categorization/  # 95% accuracy, $0.001/signal
├── tests/                # 97 tests for model validation
└── docs/                 # Deep-dive guides

Navigate by Experience:

• 🟢 Beginner: prompts/ basic patterns
• 🟡 Intermediate: MODEL_OPTIMIZATION_GUIDE.md
• 🔴 Advanced: examples/epic-categorization/

---

🔐 Security & Compliance

Security Scanning: Multi-layer (Bandit, Safety, pip-audit, Semgrep, TruffleHog) Dependency Updates: Automated weekly via Dependabot API Key Management: Environment variables only, never committed Vulnerability Reporting: Open GitHub issue with security label

CI/CD Status: All security scans pass. See workflows for details.

---

📈 Performance & Scalability

Proven Metrics:

• Throughput: 5,000+ signals/week in production
• Accuracy: 95% (vs 85% manual baseline)
• Cost Efficiency: $0.001/signal average (99.7% reduction)
• Latency: <2s p95 with model cascading
• Cache Hit Rate: 95% on repeat patterns

Scalability:

• Batch processing: 50-100 signals/batch for 92% cost reduction
• Prompt caching: 90% discount on cached tokens
• Model registry: LRU-cached for <1ms lookups

→ Benchmarks & Architecture

---

🤝 Support & Ownership

Team: Product Infrastructure Issues: GitHub Issues Contributions: CONTRIBUTING.md | CODE_OF_CONDUCT.md Response Time: Best effort (open source project)

Getting Help:

1. Check docs/ for guides
2. Search existing issues
3. Open new issue with reproduction steps

---

📅 Maintenance Status

Section	Update Frequency	Next Update
Model Definitions	Weekly (automated staleness check)	Continuous
Model Pricing	Monthly	Nov 2025
Model Specs	As released	Ongoing
Security Scans	Weekly (automated)	Continuous
Dependencies	Weekly (Dependabot)	Continuous
Prompts	Ad-hoc	As contributed

Model Update Process: Automated weekly staleness checks via GitHub Actions. See Model Update System for manual update procedures.

Deprecation Policy: 90-day notice for breaking changes. See CHANGELOG.md.

---

🎓 Learning Resources

Fundamentals (Start here):

1. Prompt Design Principles - Core patterns
2. Model Optimization Guide - Provider techniques
3. Cost Optimization - ROI strategies

Production Systems:

• Epic Categorization Example - End-to-end architecture
• Quality Evaluation - Testing methodology
• Meta-Prompting - Iterative improvement

Migration:

• Old → New System - Upgrade from legacy registry

→ Full Learning Path

---

🤝 Contributing

We welcome production-tested contributions with real metrics:

Quality Bar:

• ✅ Quantified results from actual usage (no toy examples)
• ✅ Before/after metrics (accuracy, cost, latency)
• ✅ Failure modes documented
• ✅ Tests included (97 existing tests for reference)

Process:

1. Review CONTRIBUTING.md
2. Sign commits with DCO: git commit -s
3. Submit PR using template

Recognition: All contributors listed in ATTRIBUTION.md.

---

📄 License

MIT License - Commercial use, modification, distribution allowed. No warranty.

TL;DR: Use freely in commercial products, no attribution required (appreciated).

Details: LICENSE | LICENSE_FAQ.md | CONTENT_LICENSE.md

---

🔗 Quick Links

Getting Started: Design Principles → Model Guide → Examples

Repository Structure: Project Structure Guide | Changelog Developer API: ai_models Package | Migration Guide CI/CD: Workflows | Run Tests Model Registry: Definitions | Pricing

---

Note: All examples genericized for public sharing. No proprietary data included. Metrics approximate and rounded for privacy.