State-Of-The-Art Prompting For AI Agents

YC partners Garry Tan, Harj Taggar, Diana Hu, and Jared Friedman break down state-of-the-art prompting for production AI agents, featuring Parahelp’s open-sourced six-page agent prompt.

• Parahelp’s production prompt is six pages, uses XML tags, markdown headers, and explicit output format specs to orchestrate multi-agent pipelines for Perplexity, Replit, and Bolt.
• Parahelp considers evals—not prompts—their crown jewels; without evals you cannot understand why a prompt was written or improve it.
• Metaprompting (feeding a prompt back to a large model to improve itself) is now standard practice; Tropier uses ‘prompt folding’ where one prompt dynamically generates specialized child prompts.
• o3 rigidly follows rubrics while Gemini 2.5 Pro reasons around exceptions, making model choice matter for judgment-heavy scoring tasks.
• YC’s internal trick: add a ‘debug_info’ output field so the LLM reports confusing or underspecified instructions as a developer to-do list in production.
• Vertical AI founders close six- and seven-figure deals by acting as forward-deployed engineers—demoing a custom-tuned agent built from the first meeting before the second meeting.
• Distilling a metaprompted large-model prompt down to a smaller fast model is a common latency fix for voice AI agents where humans detect pauses above a threshold.

2025-05-30 · Watch on YouTube