Clickmax

Clickmax is Bilhon's platform for pages, funnels, CRM and sales. Dark mode already existed, but every screen handled it on its own: there were more than 20,000 hand-written dark-mode classes in the main app alone. Fixing one screen didn't fix the next.

The team kept reporting the symptoms: white text on a white background in Products, unreadable lists in the messaging channels and 1.6 s of white screen on every load in dark mode.

Underneath it all was one cause: duplicated components. On October 2, 326 files still used the form fields that predated CXDS, the platform's design system, and 37 used local tabs. Of those files, 295 belonged to the product screens and made up the scope; the visual builders and the Kanban were left out by product decision, since they have their own roadmap. The old field was 56 px tall; the CXDS one, 40 px. Two platforms living on the same screen.

An inventory in the code. A script counts the files that still import an old field or tab. It ran on every merge, and that count, not an estimate, was the number I reported to the team.

The ladder. Proposed by André, the Design Lead, and applied to every screen: use the component that already exists; adapt the screen to an existing pattern; create a new pattern in the design system, for everyone; only then allow an exception. A serious visual bug became an urgent fix; a missing pattern became a component.

The guardrail. In 4 days without a guardrail, 5 new files were created with the old field. Before the first replacement, a lint rule started flagging every new use. From then on, the number only went down.

Area by area. Marketing, Members area, Products, Leads and Contacts, Messages, Integrations, Settings, Sales and, last, the sign-in screens. Each area became one PR, always following the same path, described in the next section.

I built it with Claude Code acting as a senior developer. The rule that made the biggest difference: it decides anything technical and reversible on its own, and asks about anything that touches product, visuals or someone else's work, or that can't be undone.

Scale. Each PR ran in an isolated copy of the repository, with a kickoff prompt that set the scope and the limits. That way, several areas moved forward at once without stepping on each other.

Speed on the repetitive parts. In the first area, a codemod replaced 70 labels, 34 text fields and 7 text areas in 4 seconds. The 31 cases that weren't mechanical were done one by one.

A second read. Before every commit, another session read only the diff, looking for bugs. One of those reads surfaced a bug that was already in production: when a subscription switched from annual to monthly, the 12 installments stayed active behind the scenes, while the field showed empty on screen.

Every component swap meant going over the screen with a fine-tooth comb. Along the way, I found problems that were already in production, in places nobody had asked me to look. These four had the biggest impact: they blocked people from using the product, hid errors from whoever was filling in a form, or affected billing.

Two-level tabs. André and I settled on a single rule: if any tab has sub-tabs, the main strip becomes pills with icons, a new component in CXDS, and the sub-tabs attach to the panel. Whatever doesn't fit the width goes into a “More” menu, which shows the page that's open.

In fix-up work, a correct screen just looks normal. The difference only shows side by side, which is why every release came with before-and-after screenshots.

The gain didn't stop at the screens I reviewed. It changed the cost of maintaining and evolving the platform.

The same approach now covers mobile. The audit covered 80 screens and states at 375 px and 20 tables on iPad, with findings grouped by root cause. The first three fixes are already live, and each one improved several screens at once.

• Before and after is the deliverable. Without the comparison, fix-up work goes unnoticed, even by the people approving it.
• A scoreboard in the code, not in a spreadsheet. The scripted count showed the number climbing before the guardrail and kept the conversation with the team grounded in facts.
• Measure back to back. A performance measurement taken hours apart produced a false +21%. Since then, before and after run in the same window.
• Swapping components is an audit. The most serious problems were on screens nobody had asked me to look at.
• AI extends the reach of people with judgment. It replaces 111 fields in 4 seconds and runs tests overnight. Deciding what's a pattern, what's an exception and what stays out is still design work.