How this Practicum was built
A meta case: the book and this site use the same method described in the book. A person sets the tasks and checks the result. AI agents do the draft work. Version: 2026-07-09.
Calling this "a book written by a neural network" misses how the work is divided. The author and the agents collaborate under the rules in Volume 2, which makes the project a useful real-world example.
What the agents do and what the person does
| Work layer | Agent (AI) | Person (author) |
|---|---|---|
| Research | finds sources, creates research cards, and checks figures | decides what to trust and what belongs in the book |
| Writing | drafts from a brief and offers wording options | owns the meaning, voice, personal stories, and final edit |
| Editing | proofreads, checks terminology, and finds repetition | makes every substantive accept or reject decision |
| Site | lays out pages, checks links, and runs the build | decides the structure, what to publish, and what to remove |
| Verification | runs builds and link inventories | reviews what readers will actually see |
The rules behind the work
These are the same rules used in the Practicum tools because the tools grew out of this work:
- Every agent has a contract (worksheet). The contract says what the agent can do alone and what requires approval. For example, no agent publishes or commits without a direct instruction from the author. A chapter draft never overwrites the main text. Versions sit side by side, and a person makes the decision.
- Verification is built in, not saved for later (evaluation set). Every site build goes through strict checks and an automated run across all routes. Every strong claim in the book needs a source. An unsupported claim is marked and does not get published. An invented number is a hard stop.
- Uncertainty goes to a person. If an agent is unsure about a fact, phrase, or meaning, it does
not smooth over the problem. It leaves a clear note for the author, just like the
[?]rule in Olga's example. - There is a rollback point. Before any substantial restructuring, the current state is saved so the team can return to it. Changes stay small, with a check after each step.
What this proves, and what it does not
It proves that the book's method works in practice. One person and a team of agents can maintain a two-volume book, a live site with dozens of tools, and a research base while every substantive decision remains human.
It does not prove that "AI will write a book for you." Drafts are cheap. Knowing what to say, keeping a consistent voice, and taking responsibility for every claim are expensive. That is the book's main argument, tested on the book itself.
See the tools the author uses in the Author's stack, compare system levels in Examples by Level, or build your own with the Personal Harness Builder. The code is on GitHub.