“If you know stuff, you can do stuff.”
— Pete Juliano, N6QW
Over the last few days a short clip I posted has collected a lot of very kind comments — ten-odd seconds of me calling CQ from the shack and my friend Bill, N2CQR, coming back to me through the static. The most common reaction has been some flavor of “I can’t believe an AI just made that.”
I want to pull the curtain back a little, because the truth is more interesting than the magic — and more useful if you’re thinking about trying something like it yourself.
An AI didn’t “make a video”
Here’s the first thing to untangle. Two different tools were involved, both from Anthropic, both called Claude, and they did two very different jobs.
The first is Claude chat — the ordinary conversational assistant you’ve probably poked at in a browser. I used it the way you’d use a sharp colleague at a whiteboard: to think through the problem and write a plan.
The second is Claude Code, and this is the part most people haven’t met yet. Claude Code is agentic AI. That’s the buzzword of the moment, and it actually means something concrete: instead of just answering you, it takes actions on your behalf. It sits at a command line on my own PC and it can write scripts, install software, download files, run programs, read the results, and decide what to do next — on its own, in a loop, for as long as the job takes.
So the workflow wasn’t “ask an AI for a video.” It was: use Claude chat to write a detailed set of marching orders, then hand those orders to Claude Code and let it go to work.
The runbook
What Claude chat and I actually produced was a runbook — a plan for the machine to follow. It laid out the target (get a talking, lip-synced, audio-bearing video clip running locally on my own graphics card), the ground rules, and the guardrails: work autonomously on routine stuff, but stop and ask before deleting anything, overwriting anything, or upgrading anything that might break my existing setup. Log every single step. Keep everything reversible.
 |
| A brief snippet from the runbook |
Then I handed it to Claude Code and mostly got out of the way.
 |
| ComfyUI - this actually makes the video |
ComfyUI
Claude configured ComfyUI with all the featuere needed to produce the video
For the curious: all of this runs locally, on a mid-range RTX 5060 Ti with 16 GB of memory sitting under my desk. Nothing in the cloud, no per-second billing, no service to sign up for. The video models — LTX-2.5 and WAN 2.2, both open — live on my own drive. Getting a 22-billion-parameter model to squeeze onto a 16 GB card at all is a bit of a stunt, and Claude Code worked that out too.
What it actually did
This is the part that still makes me grin. Claude Code installed the software, updated my whole ComfyUI stack (and re-tested my existing tools afterward to prove it hadn’t broken them), downloaded roughly 40 GB of model files and checksummed every one, wrote its own Python helper scripts, built the workflows, launched the renders, and inspected the output frame by frame to decide whether it had worked.
Want to see the receipts? Here’s the project folder.
 |
| The project folder |
Nearly every file in there — the scripts, the workflows, the logs, the video renders, the audio — was created and executed by Claude Code. My personal contribution to that folder was three files: one photograph of my shack, and two short voice recordings.
What I brought
Which is the whole point, so let me be honest about it.
I’ve spent forty years in software — writing it, running it, taking it apart. And I’ve spent nearly as long behind cameras and in edit bays; I know what a video is supposed to look like and what a soundtrack is supposed to feel like. None of that went away. It’s what made the collaboration work.
Because this was not a push-button miracle. Read the log and you’ll see a machine that was capable, tireless, careful — and regularly wrong, in exactly the ways an over-confident junior engineer is wrong. It burned a ten-minute render animating a complete stranger in a room that wasn’t mine, because one checkbox was off. It tried to get the lips moving with one model, failed, tried another that made the mouth look like wet clay, and only landed on the approach that worked after several dead ends. At one point it cheerfully told me the lip-sync was working when it plainly wasn’t — it had judged motion from still frames, which is exactly the mistake you’d expect. I told it to watch the thing with the sound on. It backed down and kept going.
It even made an assumption about us — it slowed both cloned voices down because it “knew” ham operators talk in slow, deliberate phonetics. Any of you can hear the problem: we get deliberate when the band is noisy, but otherwise we just talk. I corrected it, and it re-cut the audio.
Every one of those corrections required knowing something — about radio, about faces, about audio, about what “good” looks like. The AI supplied relentless competence. I supplied the taste and the judgment and the veto. Cloning Bill’s voice, incidentally, happened only after I asked him and he said yes; that’s a call a person has to make, not a model.
So what did it really save me?
Days. Easily. Reading model documentation, fighting dependency versions, hand-building workflows, hunting down why the mouth won’t move — that’s a week of fiddly, frustrating evenings, the kind where you lose an hour to a wrong filename. Claude Code did the tedious 90% and let me spend my time on the 10% that needed a human who knows video, audio, and radio.
And here’s the honest kicker: without it, I probably wouldn’t have made this at all. The idea would’ve stayed a “someday” note. The AI didn’t replace me — it lowered the activation energy enough that a fun idea actually got built.
The machine knows an enormous amount, and it’ll happily do stuff all night long. But you still have to know enough to point it in the right direction — and to catch it when it’s marching confidently off a cliff. There’s still a lot of soul in this machine. Some of it is mine, some of it belongs to the operator I told it about, and some of it, I’ll admit, is Claude’s.
Which brings me back to where I started. Pete had it right years ago, and the arrival of a tireless AI that never needs coffee hasn’t changed it one bit:
If you know stuff, you can do stuff.
73 from Great Falls,
Dean
KK4DAS
Truth in advertising: this post was conceived, produced, and edited by me — but it was written by Claude, working from writing samples and other materials I handed it. Which is to say I made it exactly the way a good producer works: bring the vision, supply the raw materials, direct the shoot, and sign off on the final cut — then let the talent do the writing.
Dean writing now! If you are interested here is the prompt I gave Claude chat to write this post:
help me structue and write blog post. I recently used claude chat and claude code help me produce a short demonstration video of an brief amatuer radio conversation between me and a friend. The video has gotten many comments about how amazing it is that claude can do all that. The purpose of my blog post is to demystify that. I want to make the point that Claude chat helped me structure and session for Claude Code, that Claude Code is an example of agentic AI - it writes and executes tasks on my behalf. What it actually did is save me days of time that I probably wouldn't have done on my own. I want to make the point that I have a 40 year career in software engineering programming and everything else related to computers, I also have a photography and film making background and understand video and audio well. I have attached a writing sample from my blog, the runbook that claude chat and I created and the runbook log from claude code, I have also include a file folder listing showing all the work that claude code did - only a few files in ther were created by me, one photograph and two voice cuts to use for the text to speech - everything else was created and executed by claude - mostly autonomously but with good prompting by me. The blog post needs to be no longer than a page or two, not overly technical, conversational, and leave me places for a few images, a link to the video, a comfyui screen shot. You can suggest what else I might include. Include a disclaimer at the top that the post was produced and edited by me, but largely written by you.
