Designer & Operator
See the Output on Amazon ↗
Agents do the loops. A human owns the gates.
8 titles live on Amazon KDP • more in production
A production pipeline in which AI agents research, draft, and fact-check full-length guidebook manuscripts under strict editorial and truthfulness constraints. Each book moves through a forward-only stage machine; agents run the work inside a stage, and nothing advances without my approval.
SpecifiedA book brief fixes the buyer, the spine, the chapter skeleton, and the research domains.
ResearchedCapped batches of research subagents, plus deep-research and video-understanding passes on other models.
Claims verifiedEvery fact becomes a ledger row with a source and a retrieved date, then goes through fact-check.
DraftedOutline, draft, a specificity audit, and a carry-over check.
Approved → printI approve the manuscript; an automated chain lays out, paginates, and preflights the book in about four minutes.
The loop: any numeric claim older than 28 days is flagged and sent back to fact-check before a book can advance. KDP upload and the physical proof stay human.
Guardrails, as implemented
- Numbers need primary sources. A numeric claim must cite the operator itself. Forums and videos can supply texture, never figures.
- Approval is a merge. A stage only advances through a gate commit I merge; no automated workflow has write access to main, and only one script may change a book’s stage.
- State lives in files, not chat. One manifest per book, an append-only claims ledger, a disagreements log between models, and fixed-format handoff notes make every run auditable and resumable.
- Agents are budgeted. Research runs in batches of eight to ten lean subagents with a hard cap of roughly thirty tool calls each.
What it caught
- A templated fee table. Cross-checking against operator sources exposed that 51 of 56 hotel rows on the companion site carried an identical $42.00 resort fee — placeholder data the review flagged for correction.
- An overclaim in the copy. The QA gate blocked a title whose text said every figure was checked while all 24 of its ledger claims were still unverified.
- A layout break. The same gate stopped another title for a two-page layout failure and marked it do-not-upload.
- One build, in numbers. Six research agents returned 293 facts — 192 straight from operators, 84 from listings, 17 rejected. The verification log closed at 454 rows behind a 21,700-word, 92-page book.
