Build log · June → September 2026
I started on June 1 with nothing that worked. By July 8 I had a winner. On July 28 we started building the whole operation into a system: swipe what works, tear it down, write the brief, make the ad, edit it, ship it. Today a full teardown takes 48 minutes of machine time with nobody sitting with it, and the whole thing runs for about $1,500 a month where agencies and a team cost about $20,000.
The line
Every stage keeps its own clock. These are medians across the last twelve runs — not a best case, not a demo — grouped into the seven things that are actually happening.
One week off it
Six days, two brands, one person running it. These are counts of finished work carrying a full stage record, not things started. September's entire tool bill was $1,823, and this week sits inside it.
Briefs
52
Image ad concepts, each through all six stages — teardown, spec, injection, hooks, variations, work order.
Done
70
Finished ad frames, carried through 183 edit passes to get there.
Video
4
Video ads cut and delivered. Not storyboards — finished files.
Pages
3 · 23
Three complete funnels, twenty-three built pages: landing, quiz, checkout, presell, upsells, receipt.
Ready
20
A full month of email placed across seven segments, anchored to real dates rather than dropped on a grid.
The part nobody shows you
Volume is easy to fake. What isn't is a line that declines to ship, on a stated rule, and leaves the reasoning behind for a person to overrule.
Refused
On the way to seventy good ones. A frame that drifts off the product, invents a face, or misses its reference doesn't get a vote.
Refused
Each with its reason written down. An arc that couldn't stand whole — warm-up, ask and close, inside its own window — came out rather than shipping half-built.
Waiting on a person
Forty-nine of those fifty-two concepts are sitting in review. Produced is not the same word as live, and the system never pretends it is.
The numbers
To first winner
37 days
June 1 to July 8.
Cost per acquisition
−$20
A new hero page the system built, against our own baseline.
To run the machine
~$1.5k
A month, every tool included.
Saved
~$18.8k
A month, against the people doing the same job. About $225k a year.
The timeline
Me, the product, and a lot of competitors' ads to learn from.
Five weeks in, the account had something that sold.
With my developer, in Claude Code. Every step I was doing by hand became a machine that runs it the same way every time.
Built from the same swipe-and-teardown loop. It cut cost per acquisition another $20 below baseline.
Every tool's bill read and added up, then put side by side with what the same work costs in people.
Every stage timed across twelve runs. A teardown that used to be a day's work is 48 minutes nobody has to watch.
The cost
Per month. The people side is what we pay today for email, ads, strategy, editing and creator management.
Where the $1,500 goes
September's real bills came to $1,823. Part of that was testing, which is why the run rate is lower.
| Tool | What it does | September |
|---|---|---|
| Higgsfield | AI video and image generation | $1,179 |
| Claude | Every thinking step: teardowns, briefs, copy, the build itself | $400 |
| Apify | Scraping: swipes, trends, creators, comments | $101 |
| Adplexity | Competitor ad library: what's running and for how long | $100 |
| ElevenLabs | Voice, music, sound effects | $22 |
| GPT Image | Every picture and every edit | ~$21 |
| Total | $1,823 |
Why it scales
The bulk runs
Swipes, teardowns, briefs and baseline assets are made in big runs. That's where the credits go.
The team
People work from finished components and briefs, so their adjustments cost a few credits, not a fresh generation.
Creators and clippers
They get the tools and the flows on their own accounts, and they're paid on performance. Fixed cost becomes performance plus tokens.