Lessons learnt while building with AI

Over the past year, AI has let our small team experiment with new hardware, build new software, and improve our own processes and tech stack. We migrated an old WordPress site with legacy assets into a much faster and more adaptable site that we can change through a Claude Code chat. We built a website audit tool for ourselves and are now helping clients using it. We designed a circuit board, had it fabricated, and got it fully functioning on the second attempt.

This isn't a story about doing the same work faster. It's about what became possible when you use AI as a lever. We migrated our site from WordPress, built our own website audit tool and designed a physical circuit board for the first time. In every project below, we decided what to build, set the standards and checked the results. AI did a great deal of the legwork.

Leaving WordPress

Our old site ran on WordPress for years. We'd outgrown it, but moving meant migrating 71 blog posts and 293 web addresses without losing what search engines knew about us. That's the kind of job a small studio puts off indefinitely, so we had.

This year we did it. We rebuilt the site in Next.js with Claude Code working alongside us. Every post came across as structured content and every old address was redirected to its new home. When we checked, all 293 old addresses led to a live page, and none returned an error.

We did one post manually, and from there automated the rest, including quality control checks and flags for incorrect or outdated links or missing content. The entire migration ran in an evening, all links maintained, all errors fixed.

We pushed it to live after four hours, using a second process to double check it all first. Our design, our control, but the lengthy and repetitive processes ran in the background for four hours and were perfect. It would have taken us days to do manually in WordPress, and each post would have to be checked manually to boot.

Building our own audit tool

When we review a website, much of the early work is gathering evidence: loading every page, testing load speed and performance, checking the details search engines and visitors rely on, and taking screenshots. We built a tool that does that gathering. It drives a real Chrome browser, runs the tests and turns the results into a scored, plain-English draft report.

It doesn't replace judgement. It collects the evidence, so our time goes on what the evidence means for the business. A person reviews every report, and the tool won't produce a client-ready version until that review is signed off. It's the engine behind our Website Audit, and we're preparing it to offer directly to clients.

Hardware: Jellyberry

Jellyberry is our sister project: a screenless, ambient AI companion that we're developing and promoting. It's where we've pushed AI furthest, well beyond web work.

Firmware coding and testing. The device's software is covered by automated tests. To check the tests themselves, we plant deliberate bugs and count how many get caught. Our server tests now catch 95% of them, up from 69%.

3D in Blender. We connected AI directly to Blender, so it can operate the 3D software itself. We've used that to create the product visuals for the website, build rigs for testing the LED light effects and make variations of our existing organic 3D model.

Circuit boards in EasyEDA. The same kind of connection let us design the main power distribution board, have it manufactured and test it.

This is the one where I was shortest on knowledge. It got me to a working board in two attempts.

Ian Byrne

Where it fell short

Two attempts, because the first board failed. It looked complete, but none of the copper tracks connected the main power input to the pads where the LED connectors were soldered. The design had passed on screen and failed in reality.

It's the most useful thing we learned this year. AI can take you a long way into an unfamiliar field, but someone still has to test the result against the real world. On the second board we did, and it worked.

Teaching AI how we work

Much of the value came from what we wrote down. We've built our own skills, short written playbooks the AI follows, for firmware testing and for building websites. They cover bringing up new hardware and running our test protocol, through to our design rules, our copy rules and the checks a page must pass before we call it done. The craft is ours. The skills mean the AI works to our standard rather than its own.

What's next, and the smaller jobs

We're now developing a 3D character production pipeline that turns a few photos into a printable figure, building on our years of 3D scanning and printing.

Not everything was a big project. Keeping our own software environment in shape counts too. When our main workstation's system drive ran low, AI built us an interactive map of where the space had gone and explained each item before anything was removed. We got back 61 GB, and we understood every gigabyte.

What it taught us

• Start with the problem. Then take the knowledge you have, research and deep dive on the topic, come back with an action plan, and tackle more ambitious projects with the help of AI that is carefully guided.
• Keep the judgement. AI is great for brainstorming and ideation, but the decisions and sign-off stay with you. If you are giving it a practical task, make it specific and bounded. The more detail you can provide, the better your results.
• Test against reality. The first board looked right and wasn't, but we were quite convinced it was until we wired it up and plugged it in. Don't be afraid to use one AI model to check on another in a new session. Sometimes you want new context and new AIs on a problem.
• Write down how you work. Our skills and notes are what made the results consistent. Have a note-taking system. In our repo we used a session-handoff skill. When our context got large or we finished a clear section, we would run the skill and we would get our session notes written up and an action prompt ready for the next session. It keeps things neater and allows you to break pieces down into smaller trackable problems and fixes.

Are you wondering how AI can be integrated into your process and help your business? With our AI Readiness Audit we look at how your business really runs, find where AI or a better process would actually help, and tell you plainly where it wouldn't.