admin
-
From Clever Generalist to Production Teammate: Building a Custom AI Skill Library
A general-purpose AI is a brilliant generalist that forgets your brand, your tagging rules, and your real prices the moment the chat ends. The leverage is in skills — reusable packets of your org’s knowledge and procedure the AI loads on demand. Here’s the playbook I used to turn a general model into a production…
-
The Fix Wasn’t the Feature, It Was the Front Door: Rebuilding ‘Help Me Choose’ (Part 2)
In Part 1, the quiz converted far better than the page average — and almost nobody could find it. So in the rebuild we resisted the obvious instinct to improve the quiz and fixed discovery instead: a ghost button became a primary CTA, abstract spec questions became plain-language intent paths. Promoting the front door didn’t…
-
What the AI Didn’t Know That I Did: Building Real Products With AI as a Domain Expert
AI wrote most of the code on a multi-week build. It also confidently tried to ship a cart that showed the wrong price, a sales bar nobody could see, and a screen that said the same thing four times. Every catch came down to something I knew that the model couldn’t. Here’s the back-and-forth, the…
-
I Gave an AI Agent Publishing Access to My WordPress Site. Here Are the Four Walls I Hit.
I wanted to draft in Markdown and have an AI agent publish straight to my self-hosted WordPress site — no copy-paste, no editor. The pipeline itself is simple. Getting past the host firewall and WordPress auth took four specific fixes that almost nobody documents. Here’s the build, the four walls, and how to get through…
-
We Built a Product Recommendation Quiz for VPS Hosting. Here’s What the Data Taught Us.
Most VPS shoppers aren’t sure which plan they need, so they leave. We built a 4-question “Help Me Choose” quiz to fix it. The data was a surprise — under 2% of visitors clicked it, but those who did converted at over 3x the page average. The quiz worked. The problem was nobody could find…
-
AI Wrote 90% of the Code. I Caught 100% of the Mistakes That Would Have Shipped.
Three specific, plausible-looking mistakes from building a plan-picker widget with AI — a cart link that would have shown a different price than the page, a “talk to sales” footer that wouldn’t render until I reframed the problem, and a triple redundancy the AI never flagged. The only mode that ships clean on a high-stakes…
-
AI Got to 80% on the Landing Page. The Last 20% Needed Domain Knowledge.
I had Claude build a landing page. Two versions, clean-looking, almost shipped. Then I caught an invisible-CTA specificity bug, a silently-dropped font import, a contact form that would have routed every lead to the wrong queue, and eight invented FAQs. AI gets to 80% fast. The last 20% needs domain knowledge.
-
Compressing Months Into Days: Rebuilding a Configurator With AI in a Week
Picture a configurator bleeding a third of its users at the configuration step. The old playbook to fix it — brief, design sprint, agency, six weeks, a five-figure budget. Instead I spent a week with Claude reading the live page, iterating on real production data, and debugging from the same browser that ran the prototype.…
-
SOP Compliance ≠ Analytics Correctness
I audited a new build against production tag by tag. It came back clean — except for two “violations” my standard would have me fix. I overrode both, because fixing them would have silently broken every dashboard calibrated to the old values. An audit tool tells you what doesn’t match the spec; it can’t tell…
-
I Ran 160 AI Shoppers Through Two Pricing Pages Before a Single Human Saw Either One
I redesigned a pricing page, then built a panel of 160 persona-conditioned AI agents to shop both versions blind — before spending a drop of real traffic. High-value plan selection jumped from 47.5% to 73.8% (p=0.0007, replicated across two runs), and the budget shoppers didn’t budge. Here’s how the simulation works, what it found, the…