Test your prompts with your own agent

LLM-as-judge is a solved pattern in research and enterprise tooling: generate candidates, score them against criteria, keep the winner. Anthropic describes the same shape as the evaluator-optimizer workflow — use it when you have clear criteria and iteration gives measurable gains. Prompts qualify on both counts.

But the tooling assumes a team, a CI pipeline, and an eval budget. What most individual developers actually do is ask the model to critique the prompt, paste the suggestion back, and hope. No record of what changed, no comparison against real output, no way back when the "improved" prompt turns out worse.

Here is the same evaluator-optimizer shape, sized for one person. The key move: the judge isn't a rubric in a YAML file — it's your own agent producing real output for each variant, so you compare results, not vibes.

ple loop code-review --exec 'claude -p {}' --rounds 3

# each round:
#  1. generates 2-3 variants of your prompt (concise / structured / chain-of-thought)
#  2. pipes EACH variant through your real agent — output appears under the variant
#  3. you pick the winner by looking at actual results
#  4. winner is saved as a new version with a Loop(strategy) changelog

Because every winner lands in version history, the iteration is recorded: v4 caught the race condition that v3 missed, and the diff shows exactly which sentence did it. OpenAI's GPT-5 guide calls the manual flavor of this metaprompting — asking the model what to add or delete — and Google's whitepaper tells you to document your prompt attempts. The loop does the documenting for you.

Raw output comparison works for coding prompts because failures are visible. For fuzzier prompts, give the judging pass explicit teeth: run the variant's output through a scoring prompt with numeric criteria. Ours is public — copy or fork prompt-improver, which critiques a prompt against the five rules all major labs agree on and returns a rewrite plus a scorecard, instead of vague praise.

This is not a substitute for real evals if you ship an AI product with SLAs — promptfoo and friends own that territory. It is the missing middle for the rest of us: faster than tab-switching, recorded unlike chat, and cheap enough to run every time a prompt underperforms. Try it on the worst prompt in your library first; that's where one round pays for the setup.