Documentation
Documentation fails quietly. Nobody files a bug for a step that is missing, they just fail to complete the task and leave. Generated docs make this worse, because they read confidently even where they are guessing.
The work is the evidence that it held.
Model-written docs are typically strong on structure and tone and weak on truth. They invent plausible flags, describe an interface from an older version, and explain a workaround that was removed two releases ago. Nothing in the prose signals uncertainty, so the reader trusts all of it equally.
The second failure is omission: the prerequisite that everyone on the team knows and no generated page mentions, because it was never in the prompt.
Automated checks handle the mechanical layer; a human handles the rest.
A technical writer with the product in front of them catches what no rule can: the procedure that works on the author's machine, the diagram that contradicts the table below it, the sentence that is true but teaches the wrong mental model.
That reviewer is named on the page's record, with the changes they requested, so documentation gets the same traceability as code.
Contradictions
Two pages, or two sections, giving different answers to the same question.
Dead references
Removed flags, renamed endpoints, broken links and stale screenshots.
Missing prerequisites
Credentials, versions and permissions the reader needs before step one.
Unsafe instructions
Commands that would wipe data or expose a secret if followed literally.
Unmarked uncertainty
Places where the draft guessed, written in the same voice as the places it knows.
AI documentation review
The automated pass is free, needs no account and uploads nothing. Escalate to a named reviewer only where being wrong is expensive.
