Your agent just built the wrong thing. Again.

Why this exists

The expensive part of a bad agent run isn't the tokens. It's that you only find out at the end, by reading a diff, when the thing is already built.

Every free tool that fixes this writes a spec for an engineer to read. If everyone reviewing your work reads code, go use one of those — they're good and they cost nothing. This is for when the person who decides can't read code at all.

What you get

See the feature before your agent writes a line of it.

You ask for a feature. Before writing anything, your agent gives you one page: the screen as it works today and as it would work after, both running as mocks you can tap. Where the change sits in the rest of the app. What it won't do. The strongest argument against building it. And the checks that settle whether it worked.

Two minutes to read. Then you say go, or you fix the plan while fixing it is free.

• Two skills your agent picks up on its own — one writes the plain-English card, one turns it into a page you can click
• A working HTML template you fill in, dark mode and phone layout already handled
• A real finished example from a shipped app — open it, click both mocks, take it apart
• The seven-part card structure, including the part everyone skips: the case against building it
• Install notes for Claude Code, plus how to run it as a plain prompt anywhere else