Your work doesn't need more docs. It needs a map.
The Cartographer walks a body of work (a repo, a runtime, a delivery folder) and leaves a map a stranger can use on day one. Even when that stranger is a model with zero context. Catalog, one card, stop. The whole tree never loads.
A cold model can't eat the tree.
Hand a fresh session your whole folder and it drowns. It reads everything. It understands the wrong things. Then it edits the file with the familiar name instead of the file that matters.
Docs don't fix this. Docs describe intent, then rot. They never say the three things a newcomer actually needs: what a change here touches, what it looks like it touches but doesn't, and which names in your tree are ghosts: things whose name survives but whose wiring is dead.
47 results. 0 of them are the file that matters.
That isn't a search problem. It's a map problem.
Catalog. One card. Stop.
One small file
Every noun in the territory, one line each: live, leftover, or ghost. Load it, pick your card, stop. The tree stays closed.
The trap on the door
Each card names what a change actually touches, and what it looks like it touches but doesn't. The mistake you were about to make is written where you'd make it.
Dead names, flagged
Dead services with living names get their own cards. Without one, a reader trusts the name and "re-enables" something that was shut off for a reason.
A change traces in two hops: catalog → card → done. You never load the queues, the gates, or the ghosts.
This is a ghost card, verbatim.
# screen-capture status: ghost source: ops/screen-audit/capture.sh ; LaunchAgent plist com.fleet.screen-audit what: A periodic screen-capture service whose NAME survives in the tree and in older notes, but whose wiring is down: the LaunchAgent was unloaded after it captured sensitive session content without erroring (the OS returns exit 0 even when the permission is missing). looked-for: `launchctl list | grep screen-audit` on each box (not loaded); recent frames in ops/screen-audit/frames/ (none newer than the incident). hits: nothing at runtime — that is the point of the card. does-not-hit: your change plans. A reader who trusts the name will "re-enable monitoring" and re-open a credential-exposure hole. If it comes back, it comes back through the self-disabling variant in capture.sh, not the plist.
One noun. ~120 words. A cold session reads this in two hops and never steps on the landmine. Names and paths are aliased; the incident is real.
The repo opens August 24.
Not a launch funnel. A date. The method, the templates, and the worked example go public that morning. The waitlist just gets the link first.
Be there when it opens.
One email when the repo unlocks. Nothing before, nothing after. If you map systems for a living, or you hand work to models that walk in cold, this is built for you.