Side projects on top of a day job: a dozen shipped apps — web, macOS, even one on the Mac App Store — ten agent skills, and two publications that write themselves daily. Built between meetings, after the dog falls asleep, mostly on Sunday afternoons. None of it is theoretical, and most of it has broken in production at least once.
I get asked the same question in every meeting now: "But how do you actually use this stuff?" Easier to point at a page than re-explain.
The average AI advice online is a thread that says "prompt engineering is the new programming" and then sells you a course. I'd rather give you specifics.
If a non-engineer can ship apps people actually use, that's signal. Take it as one. The bar moved.
Click through if you want to break them. Screenshots are the actual deployed apps.
A digital audio workstation that runs in your browser tab. Multitrack clips, synth and drum instruments, a mixer, audio recording, autotune, automation lanes, WAV and stem export. Raw Web Audio, no framework, no plugins to install.
A three-oscillator synthesizer you can play with your keyboard. Filters, envelopes, LFOs, an effects rack, dozens of presets, MIDI support — and every patch serializes into the URL, so sharing a sound is sharing a link.
The one that made it to the Mac App Store. A per-app audio mixer for macOS that lives on your letter keys — tap a shortcut, ride every app's volume in real time. Paid product, sandboxed, real reviews from real strangers.
The one with a domain of its own. A multi-tenant analytics, pacing and creative-ops platform for performance agencies that run their whole stack on data — magic-link sign-in, per-tenant isolation, edge API.
A Win98 desktop portfolio. Deliberately. My CV runs in a fake operating system because the world has enough purple-gradient landing pages and I am not adding to it.
A cozy multiplayer social farming game on Cloudflare Workers, Durable Objects and D1. Built because I wanted to learn DOs and "build something fun" beats "build a TODO app." Your farm grows in real time, even when you sleep.
A FIFO P&L portfolio tracker I actually use. Existing apps either hid the math or charged $9/month for it. Hono + D1 + KV; the dashboard tells me the truth.
A JARVIS-style assistant living in my menu bar. On-device ears (Whisper), an agent for a mind with persistent memory, a British voice, and a waveform overlay that reacts while she listens, thinks and speaks. I talk to my computer now and it's normal.
I own a physical volume mixer whose software didn't do what I wanted. So the agent and I watched its traffic, worked out the protocol clean-room, and built a native macOS app with the studio-panel UI it deserved. The knobs click. It sparks joy.
Maps and curated picks for trips I'd actually planned, distilled from too many bookmarks. Built before I left, useful on the ground, kept after. Private behind access — happy to demo on request.
In the lab right now: a desktop pixel pet that evolves from how you raise it, and remote control for agent sessions from a phone — or a watch.
Not everything I ship is an app. Some of it is systems that run daily with no human in the loop.
Each one started as a thing I kept re-explaining to the agent until I got tired and saved it. Hover any tile to read what it does.
If you're a non-engineer trying to ship something real with AI, these are the lessons I keep paying for.
Most failure isn't the model being dumb. It's goal drift — you didn't say what success was, so the agent picks one for you and gets 80% of the way there before you realize it's the wrong thing. Write the acceptance criteria first. Make it a sentence, not a paragraph.
The model has seen a million landing pages and will average them. Without references you get the median: Inter font, centered hero, three feature cards, gradient on a button. Show it three sites you want to steal from before it writes a line of CSS. Output gets 10× better.
Don't paste your whole brain into every prompt. Set up files the agent reads automatically — preferences, project facts, what you tried last week. The session that starts knowing who you are is the one that ships.
A skill is a workflow with judgement baked in — gates, references, success checks. Vibes are when you re-prompt the model to "be smarter please." Build the skill once. Future you doesn't remember what worked last time.
Five minutes reading the existing code is worth two hours of regret. The agent will happily duplicate a function that already exists three folders over because it didn't look. Make it look.
Don't deploy Friday night. Your future self has plans. This isn't AI advice — it's just advice — but it applies more once you're shipping faster than you can think.
Behind all the work above.