Last time I got my tools in order — that was the Aqua Voice story. So now the tools were ready, and it was time to actually make something with AI.
The very first thing I attempted was a tool that would semi-automatically edit the pile of fishing videos I had let build up. A plain office clerk who has never done any programming, building “software” together with an AI. It probably sounds reckless. But once I tried it, I found out all sorts of things. The wins and the stumbles, honestly.
Why a video editing tool, of all things
I love fishing, and I post fishing videos on YouTube. But the editing is brutal.
You take a long piece of raw footage, hunt through it for the moment a fish hits, cut out everything you do not need, stitch what is left together, add subtitles. Finishing a single video eats hours. Shooting is easier than editing — anyone who makes videos will nod along to that.
So I thought: “Could I get AI to take over part of this tedious job?”
Concretely, here is what I wanted. Out of a long video, have the machine automatically find the moments where the sound spikes — the candidates for “a fish just got caught.” When you land a fish your voice gets loud, almost every time. So volume ought to work as a clue. As long as it hands me the candidates, I can pick the ones I want myself. Think of a treasure hunt: I wanted the AI to do the part where you work out roughly where to dig.
Claude Code, the tool that lets AI build things
This is where Claude Code comes in. Think of it as a setup for getting AI to write the actual program — the insides of a piece of software. I describe what I want in plain words, and the AI turns that into something that runs.
What a time to be alive. I work with an AI partner I call Kuro (Claude), and what I did was write up a plain-language “instruction sheet” for what I wanted to build and hand it over. Something along the lines of: “From a long video, automatically find the top 5 loudest scenes and list them for me. I will decide which ones actually get used.”
I tripped up immediately
Here is the honest part. It did not work on the first go.
I ran the tool that came back, and the one feature that mattered — automatically finding the exciting moments — was not in there yet. The parts for cutting by hand and adding subtitles were done. The automatic extraction, the whole reason I started, was missing.
But this is where I saw something genuinely interesting about AI. Claude Code told me so itself, without me having to catch it out.
That feature is not included yet. This is as far as it currently goes. To add that function, here is what would need to be built on top.
No papering over it. And for me that was an enormous relief.
Building things with AI is not something you finish in one pass. It is run it first, see what is missing, ask for the missing part — over and over, growing the thing bit by bit. That finally landed for me. Human work is not perfect on the first attempt either, is it. Exactly the same thing.
From there I went back and said, “Add the feature using a top 5 by volume approach,” and it slowly took shape. In the end it got as far as cutting out the exciting moments — starting from slightly before the peak, for a few seconds each.
The big thing I learned from doing it
Building this tool taught me one large lesson: AI has things it is good at and things it is not.
Finding the exciting bits from the volume level is work that is decided by logic, and the AI was good at it. But video editing has another side to it. “Which scene, joined to which scene, in what way, feels good?” That instinctive part — my own gut, built up over a long time making fishing videos — still does a better job. I realised I should not force all of it onto the AI. That part is better done by me.
So the tool ended up not being a “does everything automatically” machine. It settled into a division of labour: the tedious prep work — hunting for the highlights — goes to the AI, and the finishing touches that decide the flavour stay with me. It is a lot like cooking, really. You hand off the chopping and the parboiling, and you do the final seasoning yourself.
If you are about to start with AI
If you are sitting there thinking “I would love to build something with AI, but I cannot program,” there is something I want to tell you.
You can build. At the very least, you can have a go.
I cannot write a single line of code. Even so, as long as I can explain what I want in words, the AI gives it a shape. And it does not have to be perfect the first time. “Run it first, then fix what is missing” — just knowing that this is how it goes drops the barrier to making things enormously.
And one more thing. Work out which parts to hand to the AI and which parts to keep for yourself. That, I learned from this whole episode, is the knack of getting on well with AI. Try to make it do everything and it goes worse, not better. Let it help where it is strong, and put your own strengths to work yourself. That is what leads to the best result.
Next time: the story of how my blog, after six months of neglect, suddenly went completely blank one day — and how I brought it back with AI. Another raw record of tripping up and digging out.
If you missed the setup that led into this one, it is in why I chose Aqua Voice for voice input, and the blank-blog story is here: restoring a neglected blog with AI.
