← How I use Claude
Rules I work by
When something goes wrong twice, it becomes a rule. These are the ones that show how I work. Claude reads them at the start of every session.
Design stays mine
A playground, not options
I don't pick from layouts Claude made. The real type, colors, copy and components go on a canvas I can move and edit, and Claude builds what I made.
Argue, don't build
Asked for something, Claude gives a recommendation, the reasons, and what it would reject. I make the call.
My answer is final
Later critique improves my choice. It never swaps it for a different one.
Copy before layout
Words are settled before anything visual, because layout follows content.
Ban the tells
Generic AI design has a look: a blurred sticky nav, pill tags, fade-up on scroll, a tiny label above every headline. None of it without a reason.
Proof, not promises
A gesture isn't fixed until I feel it on a real phone
A test script is not a thumb.
Nothing is called fixed without proof
A picture if the fix is visible, the output if it isn't. A check that cannot fail is not a check.
The agent that grades is never the agent that built
The tester sees the spec and the running app, never the code.
Building
Quality over the quick fix
For anything other people see. If only the fast fix fits, Claude says what it costs and I choose.
Don't delete what you can't explain
Find out why something is there first.
Copy it twice before you abstract it
Wait for the third case to show the real shape.
Write it down the second time you say it
The first time is a preference.
Some of these came from others and I adopted them: Kun Chen (a reviewer that never saw the reasoning, end-to-end evidence), Sandi Metz (duplication before abstraction), G. K. Chesterton (the fence), Google and Linear (writing rules down).