$19,701

The one claim nobody else in this market will make

The median book in my catalogue has earned $17.90. In its entire life.

Two thirds of it has never cracked $25. Ninety-two percent has never reached $100.

Every course in this category sells you a screenshot. I am selling you the postmortem. You cannot be lied to by a number that is worse than you hoped, which is the only reason to trust any of the rest of it.

Here is the actual data, from my own royalty reports, pulled 15 July 2026.

The finding that costs the most to learn

The pitch you have heard is that a catalogue snowballs: books cross-promote, the algorithm learns you, and revenue curves upward while your book count grows in a straight line.

I grew a catalogue nearly four times over and measured it. Revenue per book did not move. It sat around $3.25 a month, went nowhere, and is still going nowhere.

There is no snowball. There is addition. Which means the entire business is two numbers: what a book costs to make, and whether you lose the catalogue. Everything I know is downstream of that sentence.

The product is ten programs that fail your build before you ship the disaster

The nine modules are the story. The ten tools are the product. Each one is a script that turns a lesson I paid for into a check that runs, exits non-zero when it finds the thing, and stops the batch. A rule you have to remember is a rule you will forget at book 400. A rule that fails your build is not.

My best-earning title has made $1,229 across its whole life. A single cover rejection kills a listing permanently: the ASIN, the reviews, the ranking, the sales history. I lost five in two days. The first mistake one of these stops pays for the whole thing, several times over.

And the tools themselves are held to the same standard

A tool you have never watched fail is not a tool, it is a hope. So they ship with their own tests, and the tests earn their keep. In one recent night, four bugs turned up in the scripts that build the upload queue. Two of them were quietly corrupting the catalogue.

• !
  A single hardcoded path skipped 18 books and printed a confident total. One directory prefix was wrong, so the queue script walked past an entire pen name's catalogue and reported success. The count looked fine. It was blind to a fifth of the operation.
• !
  The queue omitted every bundle: the highest-priced products in the catalogue. The script only walked one folder. Forty-five bundles, the most expensive listings there are, were simply not in the list, and nothing said so.
• !
  A safety gate false-failed all 45 bundles at once. The fix was not to exempt them. It was to re-ask the real question in a stricter way, which then immediately caught a live bundle built from an unedited manuscript.
• !
  A book edited but not yet rebuilt was invisible to the queue. The tool compared the shipped file against a backup, so a finished edit that had not been rebuilt looked like no change at all. The rule that came out of it: rebuild first, then derive. The filesystem is the only source of truth that cannot lie to you.

That is the difference between a checklist and a control. A checklist trusts you. A control trusts nothing, including itself, which is exactly why you can trust it at scale.

The receipts: ten field notes, one per tool

Each module is the incident report behind one of those scripts. That is all a module is: the failure, what it cost, and the rule that became a tool. Read them if you want to understand why the checks are shaped the way they are. Run the tools either way.

• 01
  One 1KB favicon tied all three pen names back to a single owner. Three identities, three sites, three newsletters, months of separation work. Every one of them served the same icon file, so a single search query could have collapsed the entire roster into one person. Not a link and not a byline. The image in the browser tab. produced separation-audit.sh
• 02
  One pen name existed to make readers a single promise, and all ten of her books broke it. Her site guarantees in three separate places that the endings pay off. That guarantee is the only reason she exists as a separate name. Every book under it ended dark or unresolved. No quality check ever flagged this, because each book was perfectly good. It was simply the wrong book wearing the wrong name, which is a defect no proofreader is looking for. produced lane-check.py
• 03
  The editor stage never ran for months, and every check downstream reported that it had. Forty-three books went on sale as raw first-draft model output. Worse came later: four books whose audits all came back clean turned out to be hiding 17 real defects between them, including a character who dies and then attends a wedding. produced gate-runner.sh
• 04
  Five live listings died in two days to cover rejections. A cover rejection is not a request to fix something and resubmit. By the time the mail arrives the listing is already gone, and it takes the reviews and the sales rank with it. produced cover-qa.py
• 05
  A folder name was trusted as a book's identity, and a live bundle shipped with the wrong manuscript inside it. Two folders whose names differed by one dropped letter. The build read the wrong one, and nothing anywhere afterwards compared what came out against what had been asked for. produced validate-manifest.py
• 06
  An unattended run had the wrong manuscript aimed at a live, selling listing. It would have replaced a book readers were buying with a different book entirely. Thirty lines of guard code were the only thing between that and a post-mortem. produced wrong-book-guard.py
• 07
  My private photo originals sat one ordinary command away from the public internet. The deploy script publishes the whole site folder and skips a list four items long. The full uncropped source images, over a hundred files across the three names, were safe only because I thought to add the fourth line. Everything I forgot to list was already being published. produced deploy-guard.sh
• 08
  Editing a book that is already selling puts its cover back through review against today's rules. Two products that had been live and quietly earning for months were pushed back through review and rejected the next day. The art had not changed. The metadata had not changed. I had touched the manuscript for an unrelated reason, and that alone was the entire cause. produced pre-reupload-audit.sh
• 09
  Amazon suspended the entire account, and the pipeline had done it to itself. One standalone title was published three times in three days, because no path that creates a listing ever checked whether the book was already on the store, and the check it did have trusted a stale registry that did not even contain the book. Not a book taken down. Every listing under every pen name, dark at once, earnings frozen. Reinstated after a signed declaration, with a note that the next one may not be. produced dup-listing-guard.py

Module 0 is the tenth, and it is free. It is the economics, and it is the reason the other nine are about guardrails rather than about writing.

How the machine works

One AI coding assistant orchestrates a pipeline of specialist agents. This was built and is run with Claude: it schedules the agents, runs the gates, and drives the browser. Here is the shape of it.

The specialist agents

The specific model behind each agent, the routing, why each was chosen, and every agent's full operating instructions are in the kit. The kit also ships a one-page architecture reference: the whole machine, the named stack, and what you need to run it.

The two things I sell

If that is not obviously worth it to you yet, you are not ready to buy it. I would rather you came back in six months than refunded in six days.

What is actually in the kit

No box, no bonuses, no upsell after the upsell. Five operating documents, ten tools, and the specs they read. Here is the whole manifest.

What you bring

This is a kit, not a service. It does not run itself, and it does not include the things only you can supply.

What you do not get, and why

The questions I would ask if I were you