State-Of-The-Art Prompting For AI Agents

Jared shares a six-page prompt from Parahelp, whose AI handles customer support for companies including Perplexity, Replit, and Bolt. Diana describes its role definition, step-by-step plan, tool-call rules, and structured output. The prompt sets company-wide behavior, while customer-specific workflows and examples go into a developer prompt.

The hosts describe metaprompting: using a model to revise prompts based on failures, or to create a specialized prompt for a query. For difficult tasks, Jasberry adds expert examples of bugs such as N+1 queries to guide the model. They also recommend letting agents report when information is missing or instructions are unclear, so developers can fix the workflow instead of receiving a guess.

The hosts argue that evaluation data is more valuable than a prompt alone because it explains why the prompt was designed as it was and how to improve it. They recommend learning a customer’s workflow firsthand, then turning those observations into specific tests, such as whether to honor a tractor warranty after an invoice arrives. Understanding those details can give a startup an advantage.

Drawing on Palantir’s forward-deployed engineering model, Jared describes engineers observing customers’ work and building software around real processes. He says founders should do this themselves: listen, build a tailored demo, gather feedback, and return with a working solution. The hosts point to AI startups including Giga ML and HappyRobot as examples of teams using rapid iteration and customer-specific demos to win large deals.

The hosts say models can have different working styles and suit different tasks. Harj reports that, in their scoring example, o3 followed a rubric rigidly, while Gemini 2.5 Pro was more willing to reason about exceptions. They compare prompt work to both programming and managing a person: make expectations and evaluation criteria clear, then improve the process using feedback.