The One-Liner

The week AI moved into design and I quietly upgraded the machine that runs my content.

What Happened

Thursday, April 17 — Claude Design

Anthropic dropped Claude Design.

I was at my desk when it landed and spent about an hour going through everything. It's a full design surface — not just a tool for designers, but a system that can reason about layout, hierarchy, visual decisions. Adobe dropped 1.5% on the news. Figma slid 7%.

My reaction wasn't panic for designers — it was more like recognition. This is the direction of travel and it's moving faster than most people in the industry are comfortable admitting. The question isn't whether AI will change design workflows. It already has. The question is whether the people who've been using tools like Figma as their entire identity are building transferable skills underneath, or whether they've just been operating software.

For what I'm doing — building client sites, generating landing pages, shipping infographics and cover images — this is a straight capability unlock. Less friction between idea and output. I'll take it.

Monday, April 20 — Content Work

Continued the Instagram content work. The reels process from last week held up — creation is getting faster as the format becomes more familiar. I'm not going to oversell this. Engagement still isn't where I want it to be. But the volume is consistent now, which is the prerequisite for anything else.

The instinct when things aren't working is to tweak the strategy. The more useful instinct is usually to keep going at the current strategy long enough to generate real signal. I'm trying to stay in the second camp.

Monday, April 21 — Advisor Tool

Shipped the Advisor Tool to the Content Agent on the VPS.

The pattern: Haiku runs execution, Opus runs review. Fast model does the work, smart model checks it. I've been thinking about this architecture for a while and finally had a clean enough window to build it properly. The idea is that you get most of the speed and cost efficiency of the smaller model, but with a review pass from the more capable one on anything that actually matters.

It's live, it's running, and it changes the quality floor of everything the agent produces. That's the right way to think about it — not that Opus writes everything, but that nothing ships without Opus having seen it. The ceiling stays the same. The floor goes up.

Building agent infrastructure takes time to show results. Most of what I'm doing on the VPS right now is laying groundwork that won't look like much until it compounds. That's fine. I know what it's building toward.

Tuesday, April 21 — Prospect Check-In

WhatsApp from a prospect I’d been talking to. He confirmed he's moving slowly with his new dev, his developer — a specialist who's using Claude Code to work through the codebase. Progress is slower than he would like.

I offered ecom-specific optimisation help — not to cut his dev out, but to accelerate where the leverage actually is. The codebase complexity is a real constraint, but CVR and checkout experience are mostly front-of-funnel problems. You don't need to refactor the entire codebase to fix a very high checkout drop-off rate. You need to understand what's breaking the user's confidence at the decision point and remove it.

He heard that. Conversation stayed open.

What I Built

  • Advisor Tool shipped to Content Agent — Haiku executor + Opus advisor dual-model pattern
  • Agent infrastructure on VPS extended — quality floor raised across all content outputs
  • Instagram content continued under the new reels process
  • Prospect ecom optimisation offer made — positioned as acceleration, not replacement of his dev

The Honest Take

The Claude Design drop was a good reminder that I'm building in a landscape that's actively shifting underneath me. That's not new information — it's been true for two years — but it hits differently when you watch a company's stock price move 7% in a day because of a product announcement from a competitor.

The right posture here isn't to chase every release or panic every time the capability frontier moves. It's to stay close enough to the frontier that you can absorb new capabilities quickly, and far enough from hype that you're not restructuring your whole operation every time something drops. I think I'm roughly in the right place on that spectrum.

The Advisor Tool is a small thing but it represents a shift in how I'm thinking about agents. I've been treating them mostly as fast execution layers. The dual-model pattern changes that — now there's a reasoning layer built in. That's closer to how I think about good work generally: do it fast, then check it smart.

The prospect is moving slowly and I'm not pushing too hard. The relationship is more valuable than the deal. If his dev gets the site working, great. If he doesn't, he knows I'm the person who offered to help before it became urgent. That's the position I want to be in.

Next Week

Go deep on the prospect's actual numbers — GA4, checkout data, abandoned carts — and come in with something specific, not just a general offer to help.

Have a good one. Peace. Lachy