This skill turns any agent into a relentless self-improving optimizer using the “agent loop” pattern popularized by recent autoresearch work. Instead of hand-tuning a prompt or config once and walking away at “good enough,” the agent runs a tight loop: propose one change, test it against a defined metric, keep it if it beats the current best, revert it if it doesn’t, and repeat. Over dozens or hundreds of cheap iterations, the artifact climbs steadily toward a much higher quality ceiling than manual iteration ever reaches.
WHAT IT DOES
The skill takes a target you want to improve (a system prompt, an extraction pipeline, a code-review instruction, a model config, or any artifact you can evaluate), a small set of realistic test inputs, and a handful of binary yes/no quality checks. It then drives the optimization loop autonomously: one edit per round, one score per round, winners kept and losers reverted, with a running log of what changed and why.
WHY IT MATTERS
Manual iteration hits diminishing returns fast because humans get tired and stop. An agent doesn’t. The bottleneck flips from “can we run this experiment” to “do we even know what question to ask,” which means your job becomes curating hypotheses and defining good metrics rather than grinding through trial and error. Any metric you care about that is reasonably cheap to evaluate becomes fair game for automated optimization.
WHO IT’S FOR
Anyone who maintains prompts or configs they rely on repeatedly: customer-support agents, internal workflow automations, data-extraction pipelines, code-review instructions, or ML training scripts. If you’ve ever written something, gotten it to “good enough,” and moved on, this loop picks up exactly where you stopped.