The Vision Mixer Who Fired Parallels
A man in his fifties with zero coding experience reverse-engineered a proprietary broadcast file format, built himself a Mac app, and put its output on air at the FA Cup final. Here's exactly how he d
David Steer works the biggest live sporting events on Earth. He’s a vision mixer — a technical director, in American parlance — the person sitting at the switcher desk during a football match, combining cameras, replays, and graphics as the director calls the shots. This summer he’s at the World Cup, cutting the live fixtures in Monterrey before moving on to Philadelphia for a July 4th match.
He is not a programmer. His entire coding résumé is poking at BASIC on a ZX Spectrum as a kid in the UK, decades ago. He’s a tech enthusiast — the kind of person who listens to tech podcasts with, as he puts it, no skin in the game.
And yet over a few weeks this year, he built himself a suite of professional broadcast tools that solved a problem the industry’s actual vendors never bothered to solve. His story is the cleanest template we’ve seen for how a non-technical person turns a workplace annoyance into working software. Let’s break down the playbook.
Start with a real, specific, annoying problem
For years, Steer’s workflow had one splinter in it. The graphical transitions you see in live sport — those wipes with a team badge or tournament logo — get loaded into the switcher as video files. One desk he regularly works on, the Grass Valley Kahuna, requires those files to be pre-converted into a proprietary format called SWS, using a clunky vendor utility called KWatch.
KWatch is PC-only. Steer is a lifelong Mac user. So for years, the only reason he ran Windows via Parallels on his Mac was this one janky converter. When he upgraded to a new MacBook Air, he couldn’t bring himself to install Parallels on it — which meant carrying his old laptop around purely as a KWatch machine.
That’s the profile of a perfect first project: narrow, concrete, genuinely irritating, and something you understand deeply because you live it every working day. He wasn’t trying to “learn to code” or “get into AI.” He wanted one specific splinter out of his hand.
Don’t write the thousand-word prompt
When a friend on a Discord half-jokingly suggested he just ask Claude to build him a Mac version of KWatch, Steer’s opening move was almost comically low-effort — and that’s the lesson.
He uploaded the KWatch user manual as a PDF, described his situation in what he freely admits was rambling, badly-framed prose — this is my job, this software is PC-only, I want a Mac version — and asked: do you think we could build this?
No spec document. No carefully engineered prompt. No architecture diagram. Just the problem, the relevant material, and an open question. The AI came back immediately: give me a source video file and a sample of the converted output, and let’s find out.
This keeps coming up with people who succeed at this. The winning move isn’t a giant one-shot prompt or a bare one-liner — it’s a conversation. Describe the problem like you’d describe it to a smart colleague, hand over whatever documents you have, and iterate from there.
Ship the narrowest possible version first
Here’s the part that should recalibrate your expectations: the file format Steer needed had no public documentation. Nobody had reverse-engineered it online. Claude compared the source video against the converted output, inspected the raw bytes, theorized how the format worked, and built a converter.
Steer, by his own cheerful admission, still doesn’t know what a hex dump is. He didn’t need to. He knew what the result needed to be, and he could test it.
A day and a half after that first vague prompt, he had a working app that took a video file in and spat a valid SWS file out — verified on real broadcast hardware. That was the entire original ambition, achieved in under two days.
Only then did the feature creep begin, and this is the right order of operations. One proven brick at a time: a reverse converter (a colleague pointed out you sometimes arrive at a gig to find a different desk than promised), then a preview tool — built in a single morning in a Madrid hotel room between breakfast and a Champions League call time — then support for every desk format he touches, progressive and interlaced conversion, the works. Eventually it all rolled into one app he named Macuna, a nod to the Kahuna desk that started it all.
At one point he asked for a viewer showing video and alpha channel side by side, and Claude — noticing a spare quadrant in the layout — added an audio level meter he hadn’t asked for. It was the right call. That’s the texture of working this way: less like operating a tool, more like collaborating with an eager junior who occasionally out-thinks you.
Upgrade your tooling when the friction tells you to
Steer’s first week was pure caveman mode: chatting with Claude in the browser, pasting code into the Mac’s Terminal by hand, screenshotting errors, pasting them back. When his context window filled up, he’d ask for a handover document to seed the next conversation.
It worked. That’s worth stressing — you can get a real, hardware-tested application shipped from the chat window alone. But when his developer friends saw his workflow, they were horrified and pushed him toward Claude Code, the terminal-based version that reads and writes files directly on your machine. (In a detail too good to invent, Claude itself initially resisted the move, insisting the chat workflow was going great.)
The switch turbocharged everything. No more copy-paste loop. No more handover docs — the project now maintains its own running notes file that the AI updates and re-reads automatically. And no more shrinking giant files to upload: a six-gigabyte broadcast file can just sit on the desktop, and you point the AI at it.
The meta-lesson: don’t optimize your setup on day one. Start in the dumbest workflow that works, and graduate when — and only when — the friction becomes obvious.
Let the machines check each other
Two of Steer’s habits are quietly sophisticated. He kept documentation and a user manual throughout, and periodically pasted them into ChatGPT to sense-check what Claude had written — catching gaps and errors he’d never have spotted himself.
Better still: at one point Claude in the chat window confidently theorized how one undocumented format worked. Steer pasted the theory into Claude Code, which chased it down a rabbit hole, got nowhere, abandoned it, and cracked the format a completely different way — after which the chat-window Claude conceded defeat. Steer found himself refereeing an argument between two instances of the same AI.
He’d accidentally recreated, by hand, what the professionals call agentic coding — multiple AI agents generating, checking, and challenging each other’s work, converging on an answer like a jury. You don’t need to orchestrate an agent swarm to get the benefit. Two windows and a skeptical attitude will do.
Respect the stakes
For all the “just say yes to every permission prompt” bravado — Steer describes hammering approve-approve-approve without reading, and calls it “probably a definition of foolishness” — his approach to actual risk is disciplined. Files converted through Macuna have gone out live on air, including at this year’s FA Cup final. But the one desk model he hasn’t yet verified is the one at the World Cup, and he flatly refuses to test on tournament hardware. Verification waits for downtime on a UK desk.
Two more cautions from his experience worth stealing. First, licensing: his app leans on FFmpeg, an open-source library that’s fine for personal use but has obligations if you commercialize — something the AI used without flagging, and something he’d never have known to check. If you’re tempted to sell your creation, ask what’s under the hood. Second, if you’re employed: demonstrate benefits, but don’t feed sensitive company data into consumer AI tools while you tinker.
The actual barrier is having a problem
Asked what he’d tell someone with their own workplace annoyance, Steer’s advice is almost anticlimactic: just ask. Describe what you need to achieve and ask whether it’s achievable. His total startup cost was a subscription he already had (he later upgraded tiers because he was, in his words, having too much fun to stop — his wife spent the spring as a self-described Claude widow).
The revealing anecdote is the person on his Discord who wanted to try all this and couldn’t — not for lack of skill, but for lack of a problem to solve. That’s the real gate now. Not syntax, not computer science, not budget. The people winning at this aren’t technical; they’re the ones who know exactly where their own day is broken.
You almost certainly have a KWatch in your life — some clunky, dated, works-badly-on-purpose piece of your routine everyone just tolerates. That’s your first project. Go ask the question David Steer asked: do you think we could build this?