Free skill for Claude Code & Codex
Your AI agent,
supervised.
It checks your agent’s work like a picky human would, so you don’t have to explain it twice.
Looks like a human
Opens every screen on desktop and phone, and hunts for what’s off.
Calls out AI-looking design
Clipped text, clumsy spacing, generic layouts. Named, then fixed.
Uses it like you would
Presses the buttons, fills the forms wrong, sees what happens.
Proves the code works
Runs your tests and build. Catches skipped tests and fake fixes.
“It looks great,” it said. It didn’t.
How it works
A second pair of eyes, every time.
- 01
You say it once.
It writes down what “done” means, including what you’d obviously expect but didn’t say. If something’s unclear, it asks, up to three questions.
- 02
It checks like you would.
Runs your tests and build, reads every changed line for faked work, opens every screen on desktop and phone, and uses it.
- 03
Fresh eyes judge it.
A second agent that never saw the work hunts for what’s off, AI-looking design included. What it finds gets fixed first.
- 04
You get the truth.
Done, Partly done or Not done: what’s proven, what nobody could check, and what’s your call.
What you get
The answer, not a “trust me”.
- One word up front: Done, Partly done or Not done.
- How each thing you asked for was proven.
- What a person would notice, and what was fixed.
- What nobody could check, said plainly.
- Every changed line, if you want to look.
9/9 proven
- Tests10 passed
- Faked worknone found
- Screensdesktop + phone
- Second agentagrees, 9 of 9
Not checked
Safari, Firefox, a real phone
Your call
Two old users have plain-text passwords
Without it
“Verified. The UI looks great.”
Then you find five things wrong, and explain them all.
With it
“On a phone, one thing was off. Fixed, and checked again.”
You see it once it’s right.
Questions
Isn't “good UX” just its opinion?
No. It measures every screen against published standards: readable contrast (WCAG), tap targets big enough for a thumb (Apple, Android), text size and line length, labelled fields, nothing cut off or broken. Fresh eyes judge the rest, from decades of usability research.
Does it follow everything I asked?
Every message you sent is read again at the end. Each one has to be answered by something it built and proved, or it says why not. A half-followed instruction can’t pass.
Does it slow my agent down?
A small fix adds about a minute. A real feature, a few: in our test, 5 minutes instead of 1, and it caught a crash nobody had asked about. Less than one round of feedback.
Will it touch my code?
No. It only writes its notes, in a folder kept out of git. Your agent fixes what it finds; you see the result.
What's the guard?
Two small hooks. When the agent says “done” or “looks great” without the checks, it’s sent back, twice at most, then you’re told.
Can it check work that's already done?
Yes. Say “prove it” or “check what Codex did” on any project, branch or pull request.
What if it can't check something?
It says so. Another browser, real payments, an OS it doesn’t have: listed under “not checked”, never rounded up.
Does it work with Codex?
Yes, Claude Code and Codex, on Mac, Windows and Linux. Codex asks you to approve the two hooks once.
Say it once.
It gets checked.
- 1Get it free
- 2Type claude, then hi
- 3Say yes to the guard
- Free
- Claude Code
- Codex
- Mac
- Windows
- Linux
- 2-minute install
Needs Node 18+ and git. Chrome, Edge or Brave lets it open your pages.