empirical-prompt-tuning

Empirical prompt tuning improves agent prompts by dispatching blank-slate subagents to execute target instructions, gather evidence, and drive iterative refinements to the skill sources. This ensures changes are validated against realistic scenarios.

Patching and verifying agent instructions requires dispatching fresh subagents to implement and test instructions without prior context leakage, collecting artifacts to inform precise SKILL.md updates via controlled patch cycles.

Baseline evaluation requires defining 2-3 realistic scenarios including edge cases, recording clear success criteria, and executing these scenarios against the target skill using blank-slate delegation to measure current performance.

Structured frontmatter and a baseline workflow are required to define evaluation scenarios, success criteria, and subagent delegation protocols, ensuring repeatable evidence collection and controlled patching of SKILL.md updates.

Knowledge preservation links improvements to agent behavior, related skills, and operational runbooks like AGENTS.md and wiki procedures, ensuring patched prompts maintain operational context and integration across the system.

Blank-slate subagent evaluation requires structured frontmatter and baseline workflows to function, meaning without predefined realistic scenarios and success criteria, the evidence collection and prompt patching cycle cannot execute effectively.