Content Multiplier · a recorded run
Press play. One video turns into a week of posts.
A real run of the content system I built, recorded September 29, 2026. One six-minute transcript went in. 2 minutes 53 seconds later, 43 finished pieces came out, for $1.39 in model spend.
The creator is invented. Every word the engine wrote is real.
What's real and what isn't. The creator is invented. Dana Whitcombe and her plumbing company don't exist: I wrote her transcript and her fact file so that nothing on this page belongs to a real person or a real client. Everything the engine wrote is shown as it came out, with one exception. The model made up 12 web addresses, and a made-up link can land on a stranger's page, so those were swapped for reserved example addresses. The images carry the invented show's name.
The run
What went in
What came out
0 of 43 piecesPlays the recorded timing at 12x speed. No AI runs in your browser.
What came out.
Click through it the way a client would on review day. Each format lights up when its part of the run finishes.
What the run got wrong, and what I'd do about it.
I don't ship a run I haven't read. Here's what I caught in this one. I left all of it in, because a demo that hides its mistakes tells you nothing about how the thing behaves on a bad day.
-
Rule broken
It invented a number I told it not to.
Dana never timed how long her quotes used to wait. After the first run said "two days," I wrote a rule into her voice file banning any number there. The second run still wrote it five times, including on a quote card:
A number now beats the right number in two days.
A prompt is a request. The fix is a check in code that flags the phrase before anything is queued, the same way this engine already polices word counts on cards. -
Rule held
The dash rule worked on the first try.
The first run used the long dash about 190 times. One rule in the voice file took the copy you see here to zero. Some corrections stick straight away and some don't, and the only way to know which is to read the output.
-
Made up
It made up web addresses.
Only one of the writing skills is handed the brand's real links. The others guessed, and a few guesses pointed at real platforms. That's a gap in how facts get loaded, not in the writing, and it's the next thing I'd fix: one shared fact file that every skill reads.
-
Bug
One carousel crashed.
The run wrote two carousels. The second failed while rendering and got dropped, so this week has one. The pipeline logged it and carried on instead of sinking the whole run, which is the behavior you want. It's still a bug, and it's on the list.
-
Flagged itself
It said what it couldn't do.
This demo server has no photo library and no image-model key. The engine noticed both: the hook images went out as plain avatar cards and the YouTube thumbnails were skipped, each with a warning on the release. Nobody can publish around a warning they were never shown.
-
Wrong framing
It borrowed a phrase that doesn't fit her.
The YouTube description calls this
Week 12 of building my plumbing company in public.
Dana's run that company since 2014. The line came from the skill's own template, not from her transcript, which is exactly what a human read is there to catch.
How it works.
Six writers at once. Each format is its own skill with its own rules (post length, card word limits, how a carousel is paced). They all run in parallel on the same transcript, which is why the timeline above has six bars starting together.
One voice file, sent once. Who Dana is, how she talks, and what she's allowed to claim live in a small set of documents. They go to the model as one cached block, so after the first skill reads them the rest pay a fraction of the price.
It expects the model to misbehave. A parser pulls each piece out of the reply and assumes the reply is messy. Cards over their word limit go back with the exact problem named. Anything the engine can't finish becomes a warning on the release instead of a silent gap.
Then the images. A headless browser renders cards, slides and banners from templates: light and dark cards, carousel slides at Instagram and TikTok sizes, and a PDF for LinkedIn.
- Recorded
- September 29, 2026
- Model
- Claude Opus 4.8
- Run time
- 2:53
- Input
- One transcript, 868 words
- Tokens read / written
- Read from cache
- Model spend
- Built with
- Node, Express, the Anthropic API, Puppeteer
- Status
- Runs daily for a client. The tool is theirs.
Want this pointed at your content?
Tell me how your posts get made right now, start to finish. I'll tell you which parts should stop being done by hand, and what a run like this would look like on your videos.
Book a callRead how it's built