Your stale docs are now a production incident
Every codebase has a wiki page nobody has opened since 2023. A runbook for a service that got decommissioned. A setup guide with a screenshot of a UI that doesn't exist anymore.
For years that was fine. A human would open it, feel something was off, and ask in Slack. That instinct — this smells old — was the safety net. We never wrote it down because we never had to.
Agents don't have it.
An agent reads the wiki the way it reads the code: literally, confidently, with no sense of time. It can't tell 2019 from last Tuesday. So it plans against a deprecated API, writes code for a service that died two quarters ago, and hands you a PR that looks right. Nothing errors. It just quietly goes wrong.
A wrong doc used to cost one engineer ten minutes. Now it costs the team a day of work built carefully on a lie.
We solved this once already, for code. Untested code is a liability — we agreed on that a decade ago and built CI to enforce it. Prose got a pass, because only humans read it and humans compensate. That pass has expired. Docs are runtime input now, and they deserve the same pipeline.
Date what you write. A last-verified stamp costs thirty seconds and tells the reader — human or not — how much of this to trust.
Scope the authority. "Correct for deployments. Not for local setup." Agents take you literally. Use that.
Fail the build on rot. A doc untouched for six months, describing a service that ships weekly, is a failing test. Treat it like one.
And delete more. An empty shelf beats a shelf of lies.
Your docs have a reader now. It believes every word.