Building in Public #15: Battling for Websites & Switching to React
By Peter Coppinger
I'm really pleased with where app building is now. The new 004 engine is producing apps I'm proud of. I tested it on lots of projects this week. Here's one example, the Monkeyworks IT Portal: an asset inventory, employees, ordering from an approved hardware list, activity logs and a team you can invite. See for yourself.
See the prompt both builders were given
Build a comprehensive Internal IT Asset Management and Procurement portal for "Monkeyworks.com". The application requires a clean, professional dashboard design with role-based access control and a relational database backend. ### 👥 User Roles & Permissions: - Admin: Full system access (manage users, edit hardware catalogs, oversee all assets/orders). - Manager: Can view assets, add new employees, and request/order new hardware for their team. - Editor: IT Staff who can update asset status, log patch/security checks, and update order tracking. - Viewer: General staff who can view assigned assets or general read-only directories. Include a role-switching toggle in the UI header for testing purposes during development. ### 💻 System Modules to Create: 1. IT Asset Tracker (Inventory): - Data points to track: Asset ID, Device Type (Laptop, Desktop, etc.), Model, OS Version, Serial Number, Assigned User (Employee), Current Location Status (Stationary at Office [specify office/desk location] vs. Taken Home). - Security Logging: Track the "Last Security & Patch Check Date" and an "Antivirus Status" flag (Healthy/Action Needed). - Include filtering by OS, Location, and Security Health status. 2. Employee & Procurement Management (Manager View): - Employee Directory: A view for Managers to add new employees (Name, Department, Email, Role) to trigger onboarding. - Hardware Procurement: A catalog view of an "Approved Hardware List" (e.g., "Developer MacBook Pro", "Standard Office ThinkPad", "27-inch Monitor"). - Ordering System: Allow Managers to pick an item from the approved list, assign it to a new or existing employee, and submit the order. - Order Status Tracking: A visual timeline or Kanban board showing the status of orders (Requested -> Approved -> Ordered -> Imaging/Setup -> Deployed). 3. Missing Critical Features to Include: - Lifecycle & Depreciation Tracker: Track purchase date, purchase cost, and current lifecycle status (Active, In Repair, Retired). - Audit & Activity Log: A simple table showing history (e.g., "Editor updated Asset #1045 location to 'Taken Home'" or "Manager ordered a Developer MacBook Pro"). - Low Stock / Warranty Alerts: Visual badges indicating if a device's warranty is expired or expiring within 60 days. ### 🎨 UI & UX Design Instructions: - Create a modern, clean sidebar navigation system to switch between Dashboard Analytics, Asset Inventory, Employee Directory, Procurement/Orders, and System Logs. - Use a cohesive professional color palette (slate gray, navy blues, and clean accents for statuses like green for patched and amber for pending actions). - Include an interactive charts component on the main dashboard showing total assets by OS type, percentage of patched vs unpatched devices, and a breakdown of office vs. remote hardware.


Same brief, same afternoon. Left is Vorx on engine 004, right is Lovable.
I also asked Google for a random app idea and built it on both. At first I was disappointed that we took far longer than Lovable. Then I realised Lovable hadn't built a real app. No database. No AI. By default now, Lovable gives you a preview rather than a real database, which I completely get in terms of costs, but it's frustrating if you asked for a real app. Ours was a proper app, with a database and a nice AI feature that suggests a recipe from what's left in the fridge.
The weak spot is the home page before you log in. We put all our effort into the app, and everybody judges by that first look. At the moment our home pages for apps are cookie cutter, so I'll polish them, as long as it doesn't slow the build down.
Websites disaster
Then I looked at websites, and they were an absolute disaster. I'd been so focused on app quality with the 004 engine that I hadn't tested websites.
I took five website briefs (a music portfolio, a wedding invite, a travel blog, an editorial studio and a course landing page), built each on Lovable and on Vorx, and had them scored blind.
Two of our five didn't even build. On the three that did, Lovable beat us on every one. The gap wasn't structure. It was the pictures, how specific the words were, and polish.
I tried again on a newer setup the next day and stopped it after two of the four builds errored. My note at the time: "I didn't expect it to be this bad... Let's focus on the smallest project."
The wrong engine
Even though the focus is on building great apps, the website has to be great too. So how did it get this bad?
I built some sites on Tuesday and they were so bad. I was submitting bug reports all day, only to see a throwaway comment from the Chief of Staff: "I don't even know if these were built on the 004 engine." I was like, what did you just say?
I was like, what did you just say?
I dug into it, and the sites weren't on the 004 engine at all. They were on the old 002 engine, the one that was a disaster. I'd asked the AI to remove all the previous engines, because they were causing problems (the AI getting confused and so on), and never thought about them again. Turns out it delegated the job and never followed up, and neither did I.
How frustrating. I'd spent the day sending bugs to improve the 004 engine on websites, and it wasn't even the engine building them. So bloody annoying. Most of a day wasted. I've put new rules in place now: double check everything gets done.
Relentless improvements, with blind taste tests
Anyhow, I went to battle with blind taste tests on websites against Lovable... and we now win!
The report I got on Friday night said "all 4 test sites beat Lovable (café, portfolio, SaaS, electrician)." Love reading this.
all 4 test sites beat Lovable (café, portfolio, SaaS, electrician).
I completely redid how Vorx builds websites, and now we're beating Lovable in a blind taste test and matching their speed. Man, the ups and downs of building this.
Man, the ups and downs of building this.
I started with five sites and lost. So I picked one new site, a one-page team playbook for the EOS European Conference 2026 in Birmingham (the logistics, a two-day programme, contacts and a printable version), and improved the engine round after round until the AI judges picked us every time. Lovable built it once from the prompt; I rebuilt it on Vorx from the same prompt each round. The judges saw two sites labelled A and B and never knew which builder made which.
Here's every round:
The playbook rounds · 16 October
Every round was a fresh Vorx build against the same Lovable build. Three AI judges each picked the one they would ship. Separately, an AI scorer marked both sites out of 16 for design.
| Round | Design score, Vorx vs Lovable | Judges: which would you ship? |
|---|---|---|
| Round 1 | 9 vs 11 | not asked |
| Round 2 | 9 vs 11 | Vorx 3–0 |
| Round 3 | 11 vs 9 | Vorx 3–0 |
| Round 4 | 13 vs 11 | Vorx 3–0 |
| Round 5 | 10 vs 11 | Vorx 3–0 |
| Round 6 | 11 vs 11 | Vorx 3–0 |
| Round 7 | 13 vs 12 | Vorx 3–0 |
Round 1 was scored but not put to the judges. Every round the judges did see, they picked Vorx.
So on "which would you ship?" we won the judges' vote 3 to 0 in every round they voted on, and on the design score we matched or beat Lovable in four of the seven scored rounds, including the last two. The judges liked our hero and our headlines written in the team's own voice, like "Who do I call on the day?", and called Lovable's "a tidy but generic dark dashboard".


Round 7, the second pass in a row, built from the same brief. The judges saw them only as A and B.
But when I asked which they'd rather use, Lovable won 3 to 0. Lovable packed the facts in where you can find them. We spread the same facts across big photos and marketing headlines, so you scroll further to find anything. Our page was more than twice as long.
An info site isn't a showcase
This was just something interesting I spotted while testing by hand, the kind of thing an AI won't pick up without human taste. In some cases "it's not about just being beautiful, it's more about the information." When I gave the feedback, I put it like this: "For this particular type of site, it's all about giving the team information in a beautiful easy to consume, concise way. However we are producing it like it's a full production website (images everywhere) rather than an info site. We want a gorgeous header like we have but then focus on the info."
So the builder now works out from the request whether the reader is deciding (a marketing site) or already belongs (an info site). An info site keeps the beautiful header, then puts the information first: the first section on the first phone screen, both days of the programme side by side, one row per contact, and a print version of at most three A4 pages.


The same brief, built twice by Vorx. Left is the round 7 build the judges would ship; right is the first info-site build. Better, but not there yet.
We're still polishing info sites, so this is the direction rather than the finished result.
Hopefully none of that gets lost in the switchover to React... gulp!
Switching to React
Separately, and not because of the website problems, I'm moving Vorx to React. That's a big step for me, because I'm a huge fan of Vue. I think it's far better, faster and nicer. Every Vorx app so far has been Vue.
But a few weeks back I started wondering: would the fact that Vorx only builds Vue sites and apps put people off? So I looked into the stats, and React is by far the market leader.
React vs Vue · the market
Weekly npm downloads, 10 to 16 October 2026
Sometimes you just have to roll with it.
Sometimes you just have to roll with it. About five of my AI agents are now working on the React conversion, and the rest keep improving Vue apps and websites. Only new projects will be React. So far React is costing me more per build, but the quality is the same.
Behind the scenes
A big part of building Vorx is now cost management and keeping the fleet active. A big part of cost management is token management. And a big part of token management is context management.
So I've put a load of work into the wrapper around every Claude Code session. It connects remote control automatically, and it compacts or restarts the non-core workers on its own. This week I extended it so it spots when an agent is idle between jobs with a big context, restarts it smartly and starts its loop again. I used to compact and restart sessions by hand all the time. Now I'm getting far more bang for my buck, and the fleet runs at 100%. It's funny that the Chief of Staff keeps asking should we slow down, and I keep telling it full steam ahead.
It's funny that the Chief of Staff keeps asking should we slow down, and I keep telling it full steam ahead.
Tools · Orca
I've been using iTerm as the terminal for all my Claude Code agents, and I've been very happy with it. I learned all its keyboard shortcuts.
Then I came across a video showing how any desktop software can easily be reverse engineered now, and the guy gave Orca, his terminal manager, a quick plug. It looks great. Next chance I get, I'll switch a few agents over to it and see. Will report back.
25 companies have now created an account on Workforce, the tool that manages the fleet's entire backlog, and none of them has had meaningful traction with it. For me it's the most useful tool ever: I can see the status and add work so easily while on the move. I tried to simplify the onboarding with a welcome screen and a default role, agent and mission, but that doesn't seem to have helped. I probably need to onboard a few people by hand and see where they get confused. It's not a top priority, though.
Something odd: my build engine kept refusing to invent fake data. I had to explain several times over the weeks that fake data for a mock-up of a website or app is perfectly fine.
And while working on 20 things at the same time, I still had to spend an hour on Friday night doing the accounts, moving bills to the new company credit card. Hoping to automate more of that.
A prediction: Claude will have an update in the next month that takes on Grok bot directly. It would be insane if they didn't. They have all the ingredients.
Next steps
First, I need to finish the move to React and get back to the quality level we had with Vue, with everything working just as well. I'm still working on click-to-edit for React. It's what lets you click an element in the preview so the chat knows which one you mean, the way it works on Vue today.
Then I need to decide whether to make building Vue projects an option. It would double the amount of work, but I love Vue that much I might.
I've a conference in Birmingham for Success.co the next two days, but I'll keep the fleet working round the clock. Once everything is on React and I've finished testing everything, I think I can finally invite friends and family in.
I'd happily work 24/7 on this, such is my passion.
Back to it!
Stay updated: one founder, an army of AI agents, building this in public.
Get each new diary entry by email: the wins, the failures, and the one number we're chasing. No spam, unsubscribe any time.
Shipped this week 13 Oct – 19 Oct
- Collaboration: Comment emails and @mentions
- Payments: Prices in euro include VAT
- Design: Every app gets its own favicon
- Platform: Failed builds stay visible, then tidy themselves away
Comments