One optional manual build step will eventually get skipped

One command, no memory required

If any part of a build needs somebody to remember something on the day, you don't have a build, you have a ritual that usually works. Stamp a version into the file itself, somewhere unobtrusive, so a reader emailing about a fault you fixed weeks ago tells you exactly which copy they're holding, rather than you guessing.

The deciding test is whether somebody who has never seen the project can run the one command and get byte-identical artefacts. That's a stronger bar than 'the person who built it can run it correctly,' because the person who built it already carries the context a stranger doesn't have.

What stays a human decision regardless

A model can draft a register template, run the build command, or flag when two surfaces disagree. None of that changes who answers for the result when a buyer finds a mismatch. Four things stay explicit human decisions no matter how automated the pipeline gets: whether one command genuinely rebuilds the whole asset, which file is canonically the cover and who walks the register, whether print and digital are one artefact or two, and whether print sells inside your own checkout or someone else's store.

What done looks like

A stranger can run the single build command and get byte-identical artefacts to the ones currently live. A grep for the old cover filename or the old price, run across every surface the asset touches, returns nothing.