How I Built a Drop-a-File AI Video Publishing Pipeline

Every video I publish starts the same way: I drop an MP4 and an SRT file into a folder. After that, I'm done with the mechanics.

By the time I've responded to my Slack messages, a pipeline I built in Claude Code has read the transcript, found the best short-clip moments, rendered nine vertical shorts via FFmpeg, and handed everything off to my AI Director of Marketing to schedule. The copy editor reviews the captions before anything goes live. Content hits YouTube, LinkedIn, and Facebook on a staggered schedule – roughly 18 hours between long-form posts, shorter gaps for shorts in between.

One file drop. One complete publish chain.

I built the whole thing over about two weeks. I'm not a programmer.

The problem I was solving

I kind of hate going to social media. Not the connecting part – I like that. What I hate is the mechanical grind between "I filmed something useful" and "it's live and findable."

The copy. The paste. The format-for-this-platform. The upload. The schedule. The repeat. Scheduling tools help with the last step, but they assume you've already done the extraction: clipping, titling, writing descriptions, formatting for each channel. That work is still yours.

I wanted to automate everything before the scheduling too. Not because I'm lazy (though I kind of am), but because that work doesn't need me specifically. It needs consistency, attention to detail, and the ability to follow a process. Software does those things better than I do.

What the pipeline actually does

Here's the sequence, from file drop to published content:

1. I finish the edit. In DaVinci Resolve Studio, I do light edits, generate subtitles, and export the MP4 with the SRT subtitle file.
2. I drop both files into a folder. That's the entire manual step.
3. fieldway-social picks them up. This is a custom tool I built in Claude Code. It reads the SRT transcript and asks Claude to identify the best short-clip candidates – moments with a strong opinion, a tactical tip, a contrarian hook, or a clear story payoff.
4. Claude selects and scores the clips. It returns timestamps with reasoning. FFmpeg renders each clip as a vertical short.
5. Titles and descriptions get drafted. Claude writes YouTube titles, descriptions, and hashtags. Everything lands in my Obsidian vault as draft social posts.
6. The AI Director of Marketing takes over. It reads the editorial calendar and slots the new content – long-form video, shorts, and any related blog posts – into the schedule. The rhythm: roughly 18 hours between long-form blog posts and newsletters, shorter gaps between video shorts.
7. The copy editor reviews the drafts. Before anything goes out, an agent polishes the writing.
8. Content publishes on schedule. YouTube, LinkedIn, Facebook – staggered so each piece gets traction before the next one lands.

I made one decision in that entire sequence: which folder to drop the video into, which corresponds with the YouTube playlist it'll upload into.

What I built it with

Claude Code is a coding assistant from Anthropic. You describe what you want; it writes the code. I used it to build fieldway-social over roughly two weeks – no programming background, just a clear picture of the workflow I wanted and enough patience to iterate.

This is the identity reversal part that matters if you're reading this as someone who doesn't code: I haven't written a line of code that an AI didn't help me write. But I can describe a workflow clearly, and that turns out to be most of what you need.

What I can't do yet is build something indefinitely complex. But for a pipeline with well-defined inputs and outputs – video file in, published content out – Claude Code is more than sufficient.

The philosophy behind it

I want to be clear about what this pipeline is and isn't for.

It isn't designed to take me out of the work. I still show up on camera. I still decide what to say and why. I still choose which videos to make. That judgment work isn't automatable, and it shouldn't be.

What the pipeline automates is the extraction and distribution layer – the work that happens after the judgment call is made. The clipping. The titling. The cross-platform formatting. The scheduling. That work is necessary, but it doesn't require my specific combination of experience and perspective. It requires consistency, which software handles better than I do.

The goal isn't automation for its own sake. It's making sure my time goes to the things that actually need it – the content itself, and the conversations that come from it.

Perfection is the enemy of good here. The pipeline isn't perfect. It occasionally picks clips I wouldn't have chosen, or drafts a title I'd rewrite. But the alternative isn't "perfect manual publishing." The alternative is no publishing at all because the grind is too expensive. Ship it. Learn from it. Do better next time.

Why this matters for your practice

I run this pipeline because it keeps Fieldway's own content operations sustainable. I also run it because it demonstrates, directly, the philosophy behind what Fieldway builds for clients.

The underlying principle is the same whether you're distributing video or synthesizing client documents: most of the work eating your time isn't the work that requires your judgment. It's the extraction and organization that feeds your judgment work. If you can move that layer somewhere else – to a well-designed pipeline – you get your time back for the part that actually needs you.

One file drop. Everything else runs.

If you're working on a similar problem in your practice – content distribution, document synthesis, intake, or anything where the mechanical layer is crowding out the judgment work – email me. That's exactly what Fieldway does.

---

Related: How I Built an AI Publishing Pipeline | Your Pipeline Is Healthy. Your Calendar Is Lying to You.