make-it-betterAn Open-Source Skill That Improves Websites in Scored Rounds

What it is

Telling an AI to 'make it better' over and over tends to bloat a site until it breaks. make-it-better keeps the repetition and adds guardrails. Each round makes one focused change on its own branch, studies outside references first, and is allowed to change nothing or to remove things. A separate judge scores every version, so the agent that made a change never grades it. Scores combine Lighthouse with that judge. At the end the user gets a local review page with a button for every version, desktop and phone views, and a note on what changed and why, plus a branch with every version as its own commit and a '-best' branch at the top score. It never invents facts such as prices, reviews or customer names.

The problem

Repeated 'make it better' prompts make a model add code it does not need, never check other sites, and grade its own work. The result gets longer, not better.

What I built

One change per round, a separate judge, outside references before each change, an option to stop early at a target score, and a rule against invented facts. Every round is a commit, so any version can be restored.

What happened

Published as a public MIT-licensed repository with a companion article on the NexAI blog.

What it changed

• Turns open-ended AI edits into measured rounds that can be compared and rolled back

Who it fits

• Small-business owners improving a site built with an AI builder
• Developers who want measured, reversible improvement rounds