You have an AI agent subscription with a quota reset looming, plenty of tokens left to burn, and zero remaining mental capacity.

What do you do? You ask the agent, of course: “What can be improved?”

It will happily tell you about edge cases, obscure standards, and things you didn’t know existed. If you think an idea makes sense, you tell the agent to do it. And then you repeat—it will always come up with new suggestions.

I tried this on the repo for this (quite simple) blog. After 15 commits or so, Antigravity/Gemini 4 finally said to me:

“Nothing left to improve—we’ve covered every layer of the site.”

Naturally, I switched to Opus 5.5 😊 and got two more iterations out of it until it admitted:

“I don’t see anything else worth changing in the site itself.”

What actually changed?

This exercise produced a flurry of commits fixing:

  • RSS feeds and llms.txt support
  • sitemap.xml exclusions
  • Spelling and grammar errors
  • Tag consistency across posts
  • Social media icon sizes
  • Dynamic copyright year handling
  • GDPR cookie consent details

All small details that probably nobody cares about.

Was it productive? Maybe.
Did I learn stuff? Yes.
Did the blog get better? Maybe marginally.