Daily practice · AI Fluency
Daily practice with AI: where it helps, where it does not, and the security mindset behind it.
This page documents how AI shows up in my daily work. I use Claude alongside Figma and VS Code to move faster on the parts of the work I have already thought through: generating layouts I already have in mind, building pages with a design system, drafting common code I would otherwise look up.
The output is usually rough on the first try. What makes it useful is how fast I can get to the next version. AI is good for execution. It is less useful for the workflow logic and the complex B2B work that fills most of my day, where the decisions depend on context AI cannot hold.
Most of my Figma sessions start with the structure already worked out on paper. AI helps me see two or three variations of that structure quickly, which is faster than building them manually. The decision about which one to use stays with me.
For projects with an established design system, AI is good at applying components consistently. The boilerplate of "header, sidebar, content area, footer with the right design tokens" is mechanical work I would otherwise do by hand.
HTML and CSS patterns, JavaScript syntax, the small things you used to find on Stack Overflow. AI gives me these faster, in context, with my own variable names already used.
The closer the work gets to real users and real edge cases, the more I take back the keyboard. AI does not hold the context of a multi-role B2B system, the existing component patterns of a five-year-old codebase, or the trade-offs that come from working with a specific dev team. Those decisions need someone who carries that context.
I also do not let AI generate the final review pass on code. Nothing ships without me reading the diff. I check what packages are pulled in, what data is logged, and whether there are small dangerous patterns: a token logged somewhere it should not be, a permissive CORS rule, a dependency I do not recognise.
The point was not to ship a product. It was to keep the muscle warm: prompt, review, adjust, prompt again, until the thing on screen matches the thing in my head. The layout, the flow, and the polish are mine. AI was the pair programmer.
Authors I work with kept hitting the same wall: they could write theory in a document, but turning it into interactive HTML meant handing it to a developer. So I built a small notation tool that lets them mark up plain text with inline tags and get clean, interactive HTML on the other side, ready to drop into a learning module. AI helped with the parser logic. The tag syntax and how it works for the author are decisions I owned.
Working in B2B means client data is never far from the surface. I treat AI tools the same way I treat any third-party service: aware of what goes in, what comes out, and what I am allowed to share. Real user data, internal screens, anything proprietary, those do not go into a prompt. Sample data, dummy users, and abstracted versions of the problem, those do.
The speed AI gives me only stays useful if I remain responsible for what ships.
