I used to keep two Windows laptops in sync with Syncthing, a git pull and a page of settings. It worked until I forgot one of the three. Now it is one skill: /cos-synchronization.

Animated walkthrough of the /cos-synchronization skill

What it does

  1. It fetches GitHub on both machines first, so both sides start from the same remote state.
  2. It reads the state of the second laptop over SSH.
  3. It copies only the difference: files fresher than 2 days, keys and secrets.
  4. It hands the agent on the other machine a note on where each key lives and how to use it.

Where the model stays out

The comparison is a Python script. The model does not diff anything, so zero LLM tokens go to comparing files. The script returns a short JSON summary. The model reads it and asks me one question: apply? I answer yes or no, and only then does anything move.

I like this split. A script is exact and free to run, and the model does the one thing a script cannot: it tells me in plain words what is about to happen and waits for my decision.

The same idea as SpykeRose

SpykeRose is my own remote-access agent for Windows machines. The idea is the same: a bridge between machines instead of a pile of sync settings. One skill, or one button, is enough.

First real run

30 keys and secrets moved, 19 were already identical, 0 conflicts. Nothing was deleted on the target.

That last line is the one I care about most. A sync that can delete is a sync I have to supervise. A sync that only adds the difference, and shows me the summary first, I can run on a busy day.

If you keep two machines and a growing pile of sync tooling, count the pieces you have to remember. The fewer there are, the more likely the sync happens.