I did not touch any of these clips by hand. Every clip on this page is the first generation of this experiment, straight out of the pipeline. Where you see "v13", that is the QA agent's work: it re-cut and re-checked the edit on its own, version after version, until it passed. No person fixed anything in between.
This is not how I would normally ship. My usual workflow runs the edit inside DaVinci Resolve through an MCP server, driven by Claude Code on Opus 5.5, and I review and adjust it there. For this experiment I kept my hands off on purpose, to see what the agents do alone.
On October 28, Mo (@atmoio, Less Bitter) posted that "anybody claiming AI is editing their videos end to end is lying." His test: a two hour livestream turned into one 20 minute long form and 5 to 10 short clips. So I gave Claude his stream with one prompt and let it run. I did not touch the edit. Below is everything it made, what it cost, and what it got wrong.
The ask
One prompt: download the stream from YouTube, edit it with the skills already in my workspace, and give back a 20 minute long form at 16:9 and 10 shorts at 9:16, so we can cut the shorts down to the best 5. Then one more instruction: go through the edit and quality check every clip.
The source was Mo's livestream "ANTHROPIC HIRING MANAGER REVEALS ALL", 1 hour 41 minutes 42 seconds.
The long form
18 minutes 52 seconds, 16:9. Cut from the full stream, shaped into one story, with every filler, restart and swear handled.
The 10 shorts
9:16, laid out like Mo's own Shorts: the screen he is reacting to on top, his camera below, his caption look. Claude ranked them and recommended five (marked).
The receipts
Every stage was logged as it ran. Claude usage is read from the plan meter, which only shows whole percents and counts my whole account, so treat the percents as close, not exact. On the $100 plan, 1% of a week is about $0.23.
| Stage | Claude time | Plan used | Dollars |
|---|---|---|---|
| Download, transcribe, first cut of the long form and 10 shorts | 45 min | under 1% | under $0.23 |
| Edit pass and quality check, v3 to v13, overnight | 66 min | 2% | $0.46 |
| A video of me reacting to the edit, built in DaVinci Resolve (coming soon) | 42 min | 1% | $0.23 |
| Total | about 2.5 hours | 3% | about $0.69 |
Paid credits: 4 ScrapeCreators credits for the video details and transcript. Nothing else paid per use.
Ran on my machine, no extra cost: Whisper for transcription, FFmpeg for cuts and renders, DaVinci Resolve Studio (already owned) for the journey timeline.
Tokens: about 421,000 written by the model, 957,000 read fresh, and 198 million read back from cache across the three stages.
Not counted: the plan itself is a flat $100 a month whether I use it or not. The dollars above are the share of it this job used, not a new bill.
The edits
I made none. Every change below was Claude checking its own work: it transcribed short slices of each render on their own, listened at every cut, and re-cut whatever failed. It went from version 1 to version 13 before I watched anything.
- 19 restarts cut ("I, I", "take home, take home")
- About 30 half-sentences and self-corrections cut
- About 35 more filler words cut after the first pass
- 28 swears muted over the word only, each checked by hearing the audio on both sides
- Captions fixed where the transcript flipped the meaning ("LeetCode" was "Elite code")
- Long form tightened from 19:49 to 18:52; every short got shorter, by up to 7 seconds
- Loudness levelled to -14 LUFS on every file
What it got wrong
When I finally watched, I caught two things the quality check missed.
- A little jump in the long form at 0:17. A cut that lands a beat early.
- A stammer in "Stop vibe coding" from 0:07 to 0:09. The checker heard the words as fine, so it passed.
Neither is fixed on this page. The fix is not a manual edit. I point the quality check at the timestamp, it learns what a stammer looks like, and the next run catches it. The rules file gets better every time.
One prompt in. A long form, 10 shorts and a quality report out, for about 70 cents of a plan I already pay for.
- Source video: Less Bitter (Mo, @atmoio), "ANTHROPIC HIRING MANAGER REVEALS ALL", livestream, October 28, 2026. All footage is Mo's.
- Mo's post on X, October 28, 2026.
- Costs: my cost ledger for this job, one row per stage, read from the Claude plan meter and the session's saved token counts.
