A follow-along build from AI Automation Base. You start with a clean machine. You end with a local pipeline that turns a niche into a printed-product draft, a vertical video about it, and a set of captions waiting for your approval before anything is published.
What you build
| Part | What it is | What it does |
|---|---|---|
| Brain | Claude Code | Builds every part of it from the fifteen prompts |
| Memory | Postgres in Docker | Every run, plus four rules the database itself refuses to break |
| Hands | Ten jobs, one stage each | Phrase, guardrails, print file, draft product, link, music, reel, captions |
| Eyes | The Review Desk | The whole post on one screen, with Approve and Reject |
| Sockets | Provider, shortener, music, publisher | Each one a demo that answers itself, so the build needs no outside account |
Where this comes from
I built a print-on-demand pipeline for myself: a niche goes in, a model writes a line worth printing, the artwork gets rendered, a product is drafted at a provider, a short link comes back, a vertical video gets assembled, and one caption per platform is written. Then it stops.
The stopping is the part worth teaching. Most automation content in this space ends with "and then it posts for you", which is how people wake up to a suspended account and a product with somebody else's trademark on it. This pack teaches the same pipeline with the stop built into the database, not into a habit.
What you get is not my code. It is fifteen prompts that build your own version, on your own machine, with your own names on everything.
What you build
By the end you have a local app you can leave running. You give it a niche, it produces the whole post, and it puts the design, the draft product, the reel and all four captions on one screen with two buttons under them. You read it. You decide.
The one thing it deliberately does not do is publish by itself. There is no scheduler, no auto mode, and no configuration flag that turns one on. The last prompt fails the build if someone adds one later.
Do you need this? Be honest first
This is for you if you already sell printed products, or are close to it, and you have hit the point where making the listing, making the video and writing the captions is the actual work. It is also for you if you want to see how a human gate gets enforced somewhere it cannot be quietly removed, because that pattern transfers to anything with money or a promise attached to it.
Close the tab if you are looking for something that runs unattended and fills your store overnight. This pack refuses to build that, and it is not an oversight.
You do not need to be a programmer. You do need to be willing to read an error out loud and paste it back into Claude Code.
What you need
- The machine. Any laptop, Windows, Mac or Linux. The pack has a setup path for each one, starting from nothing installed. A small rented server works too, and the guide covers reaching it safely over an SSH tunnel.
- The accounts. Claude Code on your own Anthropic plan, and one OpenRouter key. Start on the free models. That is the whole list.
- No print-on-demand account, no shortener, no music service and no social accounts. All four are sockets in this lesson: a small module with a demo implementation that answers by itself, so the entire pipeline builds and runs before you connect anything real.
- The time. Setup is half an hour or so. The fifteen prompts are one session each, and they are meant to be spread out rather than marathoned.
What is in the zip
print-on-demand-os/
README.md the lesson: what you build, the rule that never bends, the paths
index.html the clickable guide, with copy buttons and a progress bar
COSTS.md every cost, with the date it was read
CREDITS.md where it comes from, what was verified, what was not, terms
paths/ three clean-machine setup guides: Windows, Mac, Linux
prompts/ fifteen prompts, 00 through 14, one session each
assets/cover.png
Prompt 00 is the constitution. It writes a PROJECT.md the other fourteen refer back to, so
that when Claude Code loses the thread six sessions in, you point at a file instead of
re-explaining the project.
What was verified and what was not
Verified: the clickable guide was driven by a script in a real browser, including the tabs, the progress bar, saved progress, every copy button, and a check that each prompt shown in the browser matches its file exactly. Every text file was scanned for keys, tokens and private addresses. Two review passes by a second model, Codex, read-only: the first raised 34 findings, six of them blocking, the second raised 25 more. Every one was read and dispositioned, and the dispositions ship inside the zip so you can see what was changed and what was argued with.
Not verified: nobody has built this app from prompt 00 to prompt 14 in one sitting, and the three setup paths have not been run on three fresh machines for this release. The live side of every socket is deliberately unwired and belongs to a later class. This is a proof of concept, distributed as is, with no guarantee that any step works on your setup. When something breaks, paste the whole error into Claude Code and ask it to fix it, then post what broke in the community so the next person gets a better pack.
Read the prompt before you run it
Hand each prompt to Claude first and ask what it plans to build. Compare that to what the prompt says it should build. A prompt costs a minute to fix. A wrong build costs an afternoon.
