Empirical Prompt Tuning

The author of a prompt cannot judge its quality. Whoever wrote the text already understands the implicit context, so re-reading it from "the same head" cannot detect ambiguity. The only reliable way to find ambiguity is to dispatch a bias-free executor, score the run two-sidedly, and iterate until improvements stop.

This skill is a method, not a tool. It runs inside Claude Code with the Task tool to spawn fresh subagents. The repository's scripts/eval-skill.sh and evals/ directory mechanize the bookkeeping; this SKILL.md is the workflow.

Inspired by mizchi/skills/empirical-prompt-tuning. The original is the canonical reference; install it via apm install -g mizchi/skills/empirical-prompt-tuning if you want the upstream version verbatim.

When to use

• Right after creating or substantially revising a skill, slash command, subagent prompt, or CLAUDE.md addition in this repo.
• When a skill is about to be promoted from "draft" to "published" (i.e. before merge to main).
• When a user reports the skill behaved unexpectedly and you suspect instruction-side ambiguity rather than model failure.

When not to use:

• One-off throwaway prompts.
• Trivial wording changes (typos, link fixes).
• When you only want to check personal taste — this method scores against frozen requirements, not preferences.