The method: working with AI without losing control

An AI that writes code writes fast, defects included. Delivered without guardrails, it produces interfaces that all look alike — same fonts, same gradients, same turns of phrase — and green tests that check nothing. For a client, speed without control is no advantage: it is a risk.

Write the method, not just the code.

• A constitution per project, read at every session: a file tree that does not move, rules that override habits.
• A lessons log: every mistake paid for becomes a rule, with its case.
• An anti-slop catalogue: banned patterns, grouped into six families (colour, typography, layout, effects, content, method traps), each with its reason.
• Measure what is produced, never its description: a colour is read in the served stylesheet, an overflow is measured on screen, nothing is deduced from a class name.
• Counter-proofs: a check that cannot fail proves nothing.

Two pieces that answer each other.

• Vantora, a personal-finance desktop app designed to stay local: import of PDF statements, photos of receipts and emails, an encrypted database on the machine, a hashed PIN, an AI assistant that runs on the computer and sends nothing outside. It was the testing ground for writing the project constitution and the lessons log. No real data is shown on this site.
• "design claude", the kit that gathers the method, 21 design verbs, 49 assistant skills, tools to measure the render and references from real projects, in a single portable folder.

Because it answers the question every client asks: how can I be sure that what is delivered has been verified? Every piece in this register goes through this method.