TL;DR: One raw video in — a podcast, a talk, a phone lesson — and out comes a lightly edited YouTube upload (sitting in Private), a blog draft with schema, clip picks with captions, platform social posts, and an email. All drafts. Nothing publishes without the owner’s click. Below is exactly how we run it, with our real measured numbers — so you can do it yourself, free, without paying us anything.
The measured proof (our own Descript account, July 6, 2026)
| Test | What happened | Cost / time |
|---|---|---|
| 60-second video | Studio Sound · 0 fillers found → 0 removed (pre-edited clip — the agent doesn’t invent work) · one 5.6s gap trimmed · 1:00.2 → 0:55.6 | 9 AI credits · under 4 minutes |
| Full hour of raw video | 59:48 → 42:21 · 35 filler words removed (kept four “uh huh” replies — real answers, not filler) · 63 dead-air gaps = 15m39s cut · studio-clean sound | 49 AI credits (< $1) · 13 minutes |
That second row is the whole business case: a 40-minute podcast comes back around 29 minutes, clean, for under a dollar of compute — and every word of it becomes raw material for the article, the clips, the social posts, and the email.
What the light edit does — and refuses to do
| ✅ Does | 🚫 Doesn’t (on purpose) |
|---|---|
| Studio-clean sound (even phone audio) | No b-roll, no music, no intros or bumpers |
| Removes “um / uh / you know” conservatively | No burned-in captions, no jump-cut style, no color grading |
| Trims dead air over ~2 seconds, including the ragged start and end | No auto-thumbnails (it suggests 3 frames + overlay words; a human picks) |
| Reports receipts: what changed, before/after duration | No fixing weak content — it polishes delivery, never substance |
| Transcribes everything (25 languages) | No publishing. Ever. The owner clicks publish. |
Raw beats fancy. My weekly videos go up essentially unedited and they work, because the value is the substance. This edit exists to get a raw recording over the publish line — it is not a video editor with bells and whistles, and anyone who tells you their AI pipeline replaces an editor is selling you missed expectations.
Do it yourself — the exact recipe, $0
1. Make three folders (Google Drive or local): Knowledge Base (a brand-voice note, who your audience is, your real links) · Raw (drop recordings here) · Outputs (the agent fills this).
2. Get Descript — free. The free plan covers your first full run (100 one-time AI credits, 60 media minutes). Two honest caveats: free exports carry a Descript watermark (fine for testing, not publishing) and files cap at 1 GB. When the weekly rhythm sticks, the ~$16/month Hobbyist plan is watermark-free and includes everything this method uses: Underlord, Studio Sound, filler removal, 1080p. You never need the bigger plans for this. No Descript at all? Run the skill anyway — you still get the article, clips plan, social posts and email, and you upload the raw video. Raw works.
3. Install the skill — which is just pasting a document into Claude:
Read / copy the Content Agent skill ⤓ Download as installable skill (.zip)
Open Claude on your computer, paste the whole skill, add: “Follow this. My newest video is here [drag the file in or paste a Drive link].” That is the entire install. It will ask you 3–4 questions and then work.
4. Review three things and click publish yourself: the YouTube upload (sitting in Private), the blog draft, the email. Plus clip picks with per-platform captions and ready social posts.
5. Repeat weekly. One recording a week becomes ~52 posts, ~500 clips-worth of picks, ~150 social posts, and 52 emails a year.
The rules that make it trustworthy
The skill enforces what we call content fidelity: every claim, quote and link must come from the transcript or the owner’s own files; quotes are verbatim; links are checked live before they’re included; image captions describe only what’s literally in frame; and everything stages as a draft. When we tested the writing pipeline, our own fact-check pass caught four fabrication attempts in the first draft — an invented testimonial, a drifted detail, two unearned labels — and that’s precisely why these rules are hard-coded into the skill rather than left to vibes.
Or hand us the tech and just talk into a camera
Where this came from
We built and battle-tested this live inside Sigrun’s SOMBA program — one of our leading examples, whose members debut the Content Agent this week — and the master version above now feeds every program we run. See also: the Quick Audit, GCT (goals, content, targeting), and the run on my own footage at dennisyu.com.

