Teardown No. 129 · Podcasts

Half a Day to About an Hour for 15 Clips: How Hugo Bowne-Anderson's agents start podcast production on their own

Hugo Bowne-Anderson runs Vanishing Gradients, his podcast, Substack and wider media operation, mostly as a solo entrepreneur with some contractors. He uses agents in a shared repository, now mainly Codex, and the work starts on its own. After a guest prep call, Codex checks Circleback every five minutes and starts the prep skills when the transcript arrives. After a livestream, it waits for the Zoom recording and transcript, then finds candidate moments, cuts and captions clips, checks them and uploads drafts to YouTube. He chooses the moments and approves what goes public. In his words, "Creating 15 podcast clips used to take me half a day. Now it takes about an hour of divided attention."

Try this step

Put the work in files an agent can edit. He already had a repository called VG Content Pipeline with previous questions, drafts and YouTube descriptions. He moved from ChatGPT to an agent in Cursor that could read and edit those files, and by the time of the lesson he was "primarily using Codex."

The problem

Each episode starts with guest preparation, event setup and promotion. After the livestream, the recording, transcript and Discord discussion feed into editorial review, then an edited podcast, captioned clips and written posts. By hand he drafted questions and Eventbrite copy from memory, made artwork from a designer’s Figma templates, edited audio in Audacity and arranged it in GarageBand. Even with ChatGPT he was “doing all that copying and pasting.”

The blueprint (as he describes it)

  1. Put the work in files an agent can edit. He already had a repository called VG Content Pipeline with previous questions, drafts and YouTube descriptions. He moved from ChatGPT to an agent in Cursor that could read and edit those files, and by the time of the lesson he was “primarily using Codex.”
  2. Write the rules down. His writing rules and preferences live in AGENTS.md, with examples of previous questions and pages. One rule: “No em dashes, of course!”
  3. Turn procedures into skills. His Markdown workflows became agent skills once he added a name and description at the top. He put a review step at the end of nearly every skill: look at what happened, suggest improvements, discuss them with him, then update the skill.
  4. Start prep from the call itself. Codex checks Circleback every five minutes after a prep call, retrieves the transcript and starts the prep skills. It brings back topics, questions, possible flows, title options and thumbnail bases. He asks for ten to fifteen titles and often combines parts of a couple. He now ends prep calls by saying the proposed flow aloud to the guest, so the transcript captures it.
  5. Send the checklist before going live. An hour before the livestream, Codex sends him an editable HTML checklist.
  6. Start clips when the recording lands. After the livestream, Codex checks Zoom until the video and transcript are ready. It finds and reviews candidate moments, cuts the video with FFmpeg, adds branding and captions, checks the result and uploads a YouTube draft.
  7. Give each reviewer one job. The agent once started generating its own version of his logo, though the skill told it to use the existing one. He added sub-agent reviewers: “One checks fonts; another checks layout.”

Stated results

  • Clip time: “Creating 15 podcast clips used to take me half a day. Now it takes about an hour of divided attention.”
  • Audience at the time of the lesson: over 13,000 Substack subscribers, more than 100,000 podcast listens and downloads over the previous year, and a collective audience of around 50,000 across channels. He does not attribute these to the agent setup.
  • Clip views, subscriber growth or revenue from the setup: none stated.

Caveats

  • The post is a write-up of a Lightning Lesson with Eleanor Berger, and it promotes their paid course on building an agentic software factory, with a discount code.
  • The time saving is his own estimate. The post gives no before-and-after reach numbers.
  • He says more agents means more decisions, and that “Producing more doesn’t automatically make the working day easier.”
  • Learning from YouTube results to guide future clip choices is the step he is “developing with Codex now”. It is not running yet.
  • Handles: X @hugobowne and LinkedIn in/hugo-bowne-anderson (both linked from hugobowne.substack.com/about). YouTube @vanishinggradients is linked from the same page. No verified Bluesky, Threads, Instagram or TikTok profile was found.

FAQ

How does Hugo Bowne-Anderson make podcast clips with agents?

After a livestream, Codex waits for the Zoom recording and transcript, then finds candidate moments, cuts and captions clips, checks them and uploads YouTube drafts. He chooses which moments to use and approves what goes public.

How much time does it save?

He says 15 clips used to take half a day and now take about an hour of divided attention.

What starts the work without him?

Polling. Codex checks Circleback every five minutes after a prep call and checks Zoom after a livestream, then starts the matching skills.

How does he catch agent mistakes?

With sub-agent reviewers that each check one thing, such as fonts or layout. He added them after the agent started making its own version of his logo.

Sources


Want agents like these working your distribution? Get It Built. Get new teardowns weekly: subscribe (double opt-in).

Get one teardown a week.

Same format, a new operator every issue. Double opt-in.