tugtool logo
Journal

Video Short Authoring Pipeline

November 18, 2026 at 7:01 AM
Watch on YouTube ↗

I’ve been posting short videos about Tug. Each one shows a single detail of the app. This is the of the newer ones, about how Tug arranges cards into numbered slots. It runs about twenty-six seconds. For such short little videos, a surprising amount of work goes into them! 😅

There’s a script, a screen recording, a narration, captions, an edit, a poster, a title and a description, and each phase depends on the others. So the AI and I worked out a process: nine phases!

Phase 1: The script

I start with an idea in a sentence, and I have the AI draft a script. It really is kind of terrible at it, and I need to go back and edit just about everything so it sounds like me and says what I want. However, I find that secondary editing so much easier than starting with a blank slate, given that these days, I’m trying to figure out so many things: how Tug should work, how to develop this video pipeline, how to get re-engaged on social media, etc.

Phase 2: The cut

I don’t record my screen by hand. Over time, I have built up a series of commands in Tug whereby the AI can drive the Tug app and record the screen as it does so. This means that every Tug interaction actually is the real app going through the motions, with the key presses, the cards moving… everything. The AI then cuts the recording into a silent movie, timed to a guess at how long each line will take me to say.

After my first couple of shorts, I decided that each should begin with bold movement and activity that I call an overture, a few seconds of a big, flat, full-frame drawing of the video’s premise. Activity is in flight on frame zero. In the video for this post, six colored windows glide in and pile up into a mess, then straighten into four columns numbered 1 2 3 4. Those columns fade into the real app’s four cards, each shape landing exactly on the thing it stood for.

To be clear, I don’t mean for the movement and motion to be frenetic: my goal is for the pace to be readable, not dazzling.

Phase 3: The voice

Next, I record the narration. The AI makes me a reading page with only my words on it, in large type, one beat lit at a time. I open my recording session in Audio Hijack. It usually takes me a few takes to get a clean recording without flubs, but not too many. I then do some processing on the captured audio, using some ffmpeg commands I had the AI research to try to conform to some best practices for voice audio.1

Phase 4: The re-time

The silent cut was timed to a guess. Now that I have a real reading of the script in my voice, I have the AI go back to find the time of every word as I actually said it, move each beat of the scene to match, and cut the movie again.

Phase 5: The captions and title

The word-by-word captions are their own movie: just the captions over a transparent background, to lay on top of the footage. They’re made with Phraso, an app I wrote for this job. More on that below.

Phase 6: The build

Once I have these assets together, I build the movie in Final Cut Pro and save out the version I’ll post.

Phase 7: The poster frame

I save out a single frame from Final Cut while I have the video loaded up, and I had the AI write me a little script to size and lay out a title graphic. I position the title and save that poster graphic from Pixelmator Pro.

Phase 8: The title and description

This is typically a light edit on the script. I had the AI suggest some hashtags initially, but I know which ones I want now, so I just add them.

Phase 9: It’s a wrap!

I post to YouTube.

Phraso

In A Whirlwind First Week, I mentioned that I’d made a new app from scratch for the as-I-speak captions, and that I’d release it once I polished it up a little. It’s out now: Phraso, open source under the MIT license.

A Phraso caption card reading SAMPLE AUDIO, with the spoken word in cyan

A video goes in, and a caption movie comes out. The spoken word is colored and pops in, the way captions do in short-form video. The movie is an ordinary QuickTime file with an alpha channel, so it drops into any video editor that reads one.

Phraso is a Mac app. It transcribes the video on the Mac itself, so the video and its words never leave the machine. Every word shows up as a chip I can click to fix, and the timings hold when I do. There are preset styles, and an inspector for making my own: font, colors, borders, shadow, how many words go on a card. When I export, I drag the caption movie straight from Phraso onto the Final Cut timeline.

It needs macOS 26 and builds from source for now. There’s no built app to download yet.

More to come!

This is the same way I work on Tug itself. Decide what I want, say it precisely, check the result, go back, edit, and try to turn one-off work into a repeatable pipeline, with the AI automating parts of the process so I never do the same fiddly thing twice. It turns out that this works for video too. 🎬

There are a lot of details in Tug I’d like to show this way. More shorts to come!

Footnotes

  1. The audio processing is a chain of ffmpeg filters, run in this order. The “reference” is the narration of a short I’ve already posted, so every video sounds like the same voice in the same room.

    Stageffmpeg filterWhat it does
    Rumble cuthighpass=f=90:p=2Always on. Takes out desk thumps and pops.
    Bass shelflowshelf=f=220Cuts whatever bass the take has over the reference, up to 12 dB. Used only when the take has too much.
    Low bellequalizer=f=105Cuts what’s still too heavy at the very bottom, up to 6 dB.
    Airhighshelf=f=6500Brightens the top, only if the take is darker than the reference, and never by more than 3 dB.
    GainvolumeBrings the take to a working level.
    Denoiseafftdn=nr=10Lowers the steady background hiss by 10 dB.
    De-essdeesser=i=0.25A light touch on the S sounds.
    Stereopan / aformatPuts the mono voice in both channels.
    ↩