Key Takeaways
- A 40 minute raw recording became a finished 13 minute 4K video with graphics, captions, b-roll and music. The AI pass took about 20 minutes of runtime.
- The stack is boring on purpose: Descript for the rough cut, FFmpeg for the actual video work, Claude Code as the thing that drives both.
- It is not one prompt and magic. The format was built and trained first. Claude follows a format it already knows.
- Michael's own rule: if AI gets you to 70 or 80 percent, you are almost all the way there. That is already the win.
- For a Bay Area business, this is the difference between publishing once a week and publishing every day.
Most people talking about AI video editing are showing you a demo. This is not a demo. Michael Le at On Point Tech Solutions in Los Gatos recorded himself sitting on the couch, watching Claude Code cut the very video you end up watching, and published the whole thing unedited by hand. Roughly 30 minutes of real time, on camera, including the waiting.
That matters, because the interesting part of this workflow is not the output. It is the honesty about what the machine did and did not do.
What Actually Got Automated
The raw recording was 40 minutes long. Before Claude touched it, Descript took the first pass: cutting the ums and the buts, cleaning up the audio, trimming the dead air. That alone brought it down to 13 minutes. Nothing exotic there, and it is worth saying out loud because a lot of AI editing content quietly skips this step and takes credit for it.
Claude Code picked it up from there and added everything that makes a video look produced: the lower thirds, the burned captions, the b-roll punches, the music bed, the branded intro and the end card. Michael gave it a raw file and a prompt. It went back into Descript to find the clips it had been told to chop, ran the render, and mixed the final master at 4K 30fps.
The Stack Is Less Exciting Than You Think
Three pieces. Descript does the transcript-driven rough cut. FFmpeg does the actual cutting, scaling and encoding. Claude Code is the layer that decides what to do and writes the commands.
Michael is direct about why Claude and not something else: he is already paying for the subscription, and heavy AI tooling racks up bills fast when you are running it daily. He ran this on Opus 5 at max, which he describes as buying a little extra thinking rather than the absolute smartest option available.
The piece people miss is the operating system underneath. Claude Code was attached to folders on his machine that already contained the brand guide, the motion kit and the format rules. It was not inventing a look. It was applying one it had been trained on across previous videos.
It Is Not One Prompt, And Saying So Is The Point
In the video Michael stops and corrects the obvious assumption. It looks like a single prompt because it is a single prompt, but there is a format behind everything and the system already knows the format. His words: it does not just come like this. You can tell it to do all this, but it does not just know it.
That is the difference between a party trick and a system. The first version of this workflow produced a different, worse format. What made this run work was the accumulated context, not the cleverness of the sentence he typed.
The 70 Percent Rule
The most useful thing in the video has nothing to do with video. Michael lands on it while waiting for the render:
AI is not about getting you to 100 percent. Sometimes it does, and that is amazing when it happens. But if it gets you to 60, 70, 80 percent, you are already almost all the way there. That by itself is the benefit.
He also does not oversell the result. Watching it back he says it could be better and more interesting, but it is okay for right now. That is a healthier standard than most AI content will give you, and it is the one that actually lets you ship daily.
Why This Matters for Bay Area and San Jose Businesses
Ask any local business owner in San Jose why they are not posting more and the answer is almost never strategy. It is that editing takes an evening they do not have. Michael's estimate for a human editor on this exact video is at least an hour, probably more, since the source was 45 minutes.
Twenty minutes of unattended runtime changes the shape of the problem. You are no longer choosing between a polished video and no video. You are choosing between a good enough video today and a perfect one that never gets made. For a Bay Area service business competing on being visible in a crowded local market, that trade is not close.
And the render runs while you work. Michael makes that point plainly: for those 20 minutes he would normally be in another session doing something else entirely.
Practical Steps
- Do the rough cut first. Run your raw recording through a transcript-based editor and strip the filler before AI ever sees it. This is where the biggest time saving actually lives.
- Install the plumbing. FFmpeg is the piece that does the real work. Without it there is nothing for the AI to drive.
- Write your format down. Brand colors, fonts, where the captions sit, what the intro does. The system can only apply a standard that exists in a file somewhere.
- Start with one video and expect the first format to be wrong. Michael's earlier videos used a different one. The current format is the result of correcting it repeatedly.
- Set the bar at 80 percent and publish. Then fix what actually bothers you on the next one.
Watch the Full Build
Final Thoughts
The headline number is 20 minutes. The real story is that a system which knows your brand can do a job that used to need a person, and that getting to 80 percent automatically beats getting to 100 percent eventually.
On Point Tech Solutions is a Go High Level consultant in Los Gatos serving San Jose and the wider Bay Area. We build websites, funnels and automations, and our clients own what we build. Done for you, or taught with an SOP so your team can run it. If you want a content system like this one pointed at your business, start at optechsol.llc.