Foundation

State who the model is and what success looks like before the task.

Explicit boundaries the model must not violate.

Force a reusable structure the user can copy or parse.

Require the model to request missing critical inputs instead of fabricating.

Show 1–3 miniature input→output pairs when the pattern is hard to describe.

Using biographical personas in system prompts to influence LLM behavior, potentially affecting output length, style, and correctness depending on the model and persona.

Iteratively refine agent harness specifications using pairwise feedback from revision history to boost agent performance and data quality for training.