A Kindle pipeline that refuses to let you publish badly.
You can already get a chatbot to write you a book. That stopped being the hard part two years ago. The reason it sells nothing has almost nothing to do with the writing, and this kit is built around that.
If you have published on Kindle and it went nowhere, it was probably one of four things. All four are predictable, and none of them is prose quality.
- You picked the topic. Because it interested you, or a video said the niche was hot. Nobody checked whether real people type that phrase into Amazon's search box. Wrong topic plus flawless execution still equals zero sales.
- You bulk generated it. One prompt, forty thousand words, an hour. It repeats itself, the examples stop matching by chapter four, and it circles the subject without ever reaching the actual instructions.
- The cover was an afterthought. A buyer decides in about a fifth of a second, at thumbnail size, in a grid, next to nine competitors. You designed at full size and never looked at it small.
- You published and stopped. Placement alone earns pennies. A listing with no marketing on top is a lottery ticket with a nice cover.
Every one of those is a discipline problem, not a writing problem, which is why a better model fixes none of them. So this is not a prompt pack. It is a set of scripts that refuse to advance when a step has been skipped.
The seven gates, as mechanisms
Each gate is a script or a state machine that fails closed. Not an instruction to be careful.
Demand before production
Probes the Amazon Kindle Store autocomplete for real buyer phrases. If the source is unreachable it writes no evidence file and exits non zero, and everything downstream refuses to start. A gate that passes when it breaks is not a gate.
Pain points are the spec
Mines the 1 to 3 star reviews of current bestsellers. What existing books fail to deliver is the sharpest specification available. The outline builder exits non zero below two named pain points.
Distribution decided first
Channel chosen before anything was built, on the question of whether it already has buyers or you must import them. The reasoning and the counter case are both written down.
Structure before generation
Eight chapters by three subchapters by three sections, so 72 addressable leaves, each with a promise, a must answer question and an explicit list of what its siblings own. Empty promises are refused.
One section at a time
The core mechanism. Hands out exactly one section and refuses the next until the current one clears four named checks: repetitive, consistent, depth, voice. There is no batch mode and no flag to add one.
The cover as its own asset
Produced deliberately by an image model, then checked by a script that parses the real image file, enforces Amazon's cover geometry and estimates title legibility at thumbnail size.
Publish, then market
A dated 14 day marketing ladder, not an upload step. Publishing is where the work starts.
Gate 5 is the one that matters most, because "write well" is not a mechanism. Any fail verdict on repetitive, consistent, depth or voice sends the section back to pending and blocks the pipeline. Depth is the common killer and the hardest to notice, because machine prose reads fine sentence by sentence while never reaching the instruction. The runner also warns you when the reviewer has passed many sections and failed none, because a reviewer that never rejects is a rubber stamp, and a rubber stamp is worse than no reviewer.
What is in the box
- Six Python scripts, stdlib only, each with
--selftest. Demand probe, outline scaffold, section runner, cover check, title score, royalty model. Nothing to install. - Five Claude skills. Demand gate, section write, section review, metadata, cover brief.
- Two scheduled task prompts. A daily book build and the 14 day launch ladder.
- An agentic installer. Hand it to your own AI agent with the unzipped kit and it installs the skills, registers the scheduled tasks on the declared cadence, and proves the install by watching the demand gate actually refuse a weak niche.
- ECONOMICS.md with the arithmetic, break even, and explicit kill criteria for when to stop.
- HUMAN_STEPS.md, RISKS.md, VERIFY.md. The verify file carries real command transcripts, not illustrations.
On 7 July 2026 Amazon raised the ceiling on the 70% royalty option from 9.99 to 12.99 USD. That ceiling had not moved since 2007.
Essentially every guide, course, blog post and royalty calculator online still says 9.99. Using Amazon's published formula at 0.15 USD per MB delivery, a 1.2 MB file returns this per US sale:
| List price | 70% royalty per sale |
|---|---|
| 9.99 USD | 6.87 USD |
| 12.99 USD | 8.97 USD |
31% more royalty on the same sale for one field on the upload form. Verified against Amazon's own list price requirements page on 4 August 2026, and the included royalty_model.py --selftest reproduces Amazon's own worked example to the cent.
To be precise about what that is: it is the royalty on one sale, not a forecast and not income. Nobody can tell you how many copies your book will sell. The script deliberately ships with no default for copies per month so it cannot produce one by accident. Full working here, sources included, free.
What you have to do yourself
Five steps. About 75 minutes once, then roughly 10 minutes per book. The KDP account, the tax interview, the bank details, the pen name, and the upload. Each is listed with the reason an agent cannot do it.
The upload is the honest one. There is no KDP publishing API and there never has been, not for you and not for large publishers. Every tool advertising KDP automation is driving the browser form. Pointing a bot at a logged in Amazon account risks the account, and the account is the business. So this kit does not do that, and tells you not to either.
Founding buyers get the first book run alongside them. Email the lab and your first book goes through all seven gates with you, watching the same reviewer refuse the same failure modes it refuses for every book that comes after yours.
Gate 1 leans on an endpoint Amazon does not contract. The autocomplete source powers Amazon's search box and Amazon can change it. That is why it fails closed and why --selftest tells you in three seconds rather than after you have written a book. A manual fallback is documented.
New title uploads are capped at three per day. If your plan was volume as the strategy, the platform already priced that in.
This is a new vertical for the lab. It reuses none of the MCP tooling. It got built because the unit economics stood up, not because it was adjacent to what already existed.
No subscription and no running cost. One file, delivered by email as a download link.
Not for sale through a checkout on this site yet. We are opening the kits to the founding list first, on purpose. Leave an email below and the link comes to you directly, at this price.
No card details here and no checkout to load. One email, one download link, on the day it opens.
Get on the list
Leave your email and the download link arrives the day it opens, at the founding price, before it goes anywhere else.