AI summary by TwinMind
State-Of-The-Art Prompting For AI Agents
Y Combinator
31m 26s
AI Startup Prompt Engineering Survey
- YC surveyed 12+ AI companies on prompt engineering practices
- Collected practical tips from frontier AI startups
- Focus on real-world implementation insights
Parahelp Case Study - Customer Support AI Agent
- Parahelp powers AI customer support for Perplexity, Replit, Bolt, other top AI companies
- Open-sourced their actual production prompt (6 pages long)
- Prompt structure breakdown:
- Role definition: "You're a manager of customer service agent"
- Task specification: approve/reject tool calls for agent orchestration
- Step-by-step plan (steps 1-5)
- Output format requirements for API integration
- Markdown-style formatting with hierarchical structure
- XML tag format for better LLM parsing
- Reasoning methodology outlined
- Worked examples provided
Prompt Architecture Layers
- System prompt: high-level company API definition (Parahelp example)
- Developer prompt: customer-specific context and workflows
- User prompt: end-user inputs (not applicable for Parahelp)
- Challenge: scaling without becoming consulting company
- Need for prompt forking/merging across customers
Metaprompting Techniques
- Prompt folding: prompts dynamically generate better versions of themselves
- Classifier prompts generate specialized prompts based on previous queries
- LLMs help improve existing prompts by analyzing failure cases
- Process: feed failed examples to raw LLM with "help me make this prompt better"
- Works because LLMs understand themselves well
Advanced Prompting Patterns
- Tropier company: prompt folding for multi-stage workflow debugging
- Jasberry: automatic bug finding using expert programmer examples
- N+1 query detection through hard examples
- Meta prompt with specific error type examples
- Pattern: when task too complex for prose, use examples instead
- Similar to test-driven development for LLMs
LLM Escape Hatches
- Problem: LLMs hallucinate when forced to give formatted output without sufficient info
- Solution 1 (Tropier): explicit instruction to stop and ask when unsure
- Solution 2 (YC): debug info parameter in response format
- LLM reports confusing/underspecified information
- Creates developer to-do list from production data
- Run in production with real user data for feedback
Model Distillation Strategy
- Use large models (Claude 3.5, GPT-4) for metaprompting
- Generate refined prompts for smaller/faster models
- Deploy on distilled models (e.g., Llama) for production
- Common for voice AI agents needing low latency
- Balance: better prompts vs faster inference
Debugging and Development Tools
- Gemini 2.5 Pro thinking traces for prompt debugging
- Long context windows enable real-time reasoning observation
- Recently added thinking traces to API
- Use gemini.google.com directly with JSON file drag-drop
- YC built internal workbenches for debugging
Evaluation (Evals) as Crown Jewels
- Parahelp considers evals more valuable than prompts
- Evals explain why prompts written specific way
- Enable prompt improvement
- Require sitting with domain experts to understand workflows
- Example: tractor sales regional manager reward functions
- Codify in-person interactions into specific evals
Forward Deployed Engineering Model
- Palantir origin: send engineers (not salespeople) to customers
- Engineers sit with FBI agents, understand actual workflows
- Convert file cabinet/fax processes to clean software
- Palantir forward deployed engineers becoming successful YC founders
- Key training for AI startup founders
Modern Forward Deployed Engineering
- Founders must be forward deployed engineers of own companies
- Cannot farm out to others
- Founders need to be: technical, product people, ethnographers, designers
- Goal: "wow, never seen anything like that" demo experience
- AI enables faster iteration than traditional Palantir model
Successful Examples
- Giger ML: voice customer support, closed huge Zepto deal
- Happy Robot: 7-figure contracts with top 3 logistics brokers
- Process: first meeting → context gathering → prompt stuffing → next-day demo
- Differentiation through technically challenging demos (low latency + accuracy)
Model Personalities and Selection
- Claude: more happy, human-steerable
- Llama: needs more steering, like talking to developer
- Different models for different use cases
- O1 vs Gemini 2.5 Pro comparison:
- O1: rigid, strictly follows rubrics
- Gemini 2.5 Pro: flexible, handles exceptions well
- O1 like "soldier" (check, check, check)
- Gemini like "high agency employee" (reasoning about exceptions)
Management and Communication Analogies
- Prompt engineering feels like coding in 1995 (tools not fully there)
- Also like learning to manage people
- Need to communicate decision-making requirements clearly
- Kaizen principle: people doing the work best at improving process
- Metaprompting embodies this principle
