Empirical Prompt Tuning

A prompt's quality is not visible to the person who wrote it. What the writer reads as clear is where another agent gets stuck. Have a fresh executor run it, evaluate both sides, and iterate is the core of this skill. Do not stop before the gains plateau.

When to use

• Right after creating or heavily revising a skill, a slash command, or a task prompt
• When an agent does not behave as expected and you want to look for the cause in the instruction's ambiguity
• When hardening a high-stakes instruction (a skill used often, a prompt at the center of an automation)

When not to use:

• A one-off throwaway prompt (the evaluation costs more than it returns)
• When the aim is the writer's own taste rather than a higher success rate

Workflow