Most prompting guides tell you to "be specific" and call it a day. A recent breakdown by AI researcher Elvis Saravia, summarizing a conversation with Y Combinator founders building production AI agents, gets far more practical than that.
The core message: treat your LLM like a brand-new employee who is brilliant but has zero context about your business. The more detailed your instructions, the better the output. One YC startup, Parahelp, uses a system prompt that runs over six pages just for its customer support agent.
Here are the techniques worth paying attention to.
The "Manager" Approach and Structured Output
The most effective agent prompts share a pattern: they assign a clear role, outline the task step by step, and specify the exact output format. Parahelp uses XML-like tags (for example, <manager_verify>accept</manager_verify>) so the LLM's responses are machine-readable and consistent.
This matters because AI agents typically feed their output into other systems. A chatbot response that sounds nice but doesn't follow a structured format breaks the pipeline downstream. Markdown headers, bullet points, and explicit tags all help.
Another YC startup called Jazzberry, which builds an AI bug finder, includes specific examples of good and bad outputs directly in the prompt. This few-shot approach (showing the model what you want instead of just describing it) significantly improves reliability for edge cases.
Prompt Folding and Escape Hatches
Two techniques stand out for agent reliability. Prompt folding means designing a first-stage prompt that classifies the input, then dynamically generates a more specialized prompt for the next stage. Instead of one massive prompt trying to handle every scenario, you get a chain of focused prompts that adapt to context.
Escape hatches are just as important. Explicitly telling the model "if you don't have enough information, say 'I don't know' and ask for clarification" dramatically reduces hallucination (when the model confidently makes things up). Without this instruction, most models will guess rather than admit uncertainty.
Model Personality Differences Matter
The founders noted that different models respond differently to the same prompts. Claude tends toward more natural, human-like responses. Llama 4 often needs more explicit steering. A practical strategy: use a larger, more capable model to help write and refine your prompts (meta-prompting), then optimize those prompts for a smaller, cheaper model you run in production.
The single most underrated piece of advice in the entire breakdown: "The prompts are important, but the evaluation suite is your most valuable IP." A good set of test cases that measure prompt quality lets you iterate with confidence. Without evals, you are guessing whether changes helped or hurt.
None of these techniques are secret or novel individually. The value is seeing how production teams combine them - six-page prompts with structured output tags, escape hatches, few-shot examples, and automated evaluation - into agents that actually work reliably enough to face real users.