| name | cerebellum |
| description | Single fast-execution skill for macOS / Linux / Windows app operations.
The "cerebellum" to the LLM's "cerebrum" — the brain decides intent,
this skill executes the canonical multi-step pattern in one syscall
sequence (no LLM round-trip per step).
Inspired by the same architecture as the LocalKin robot car: the
high-level brain plans ("rename this file to X"), the cerebellum
daemon executes deterministically (`mv old new`) at near-zero
latency.
USAGE:
cerebellum "<category> <action> [args...]"
Categories:
macOS finder, notes, mail, calendar, reminders, settings,
safari, music, photos, maps, pages, numbers, keynote,
terminal, multi, web (16 cats, 478 actions)
Linux linux-files, linux-apps, linux-settings,
linux-clipboard (4 cats)
Windows windows-files, windows-apps, windows-settings,
windows-clipboard (4 cats)
Run with no args to see the full action menu.
Examples:
cerebellum "finder rename /old.txt /new.txt"
cerebellum "finder set_view list"
cerebellum "finder set_sort Date"
cerebellum "notes create 'Title' 'body content'"
cerebellum "notes pin 'KinBench Pinned 164'"
cerebellum "notes export_pdf 'My Note' /Users/me/out.pdf"
cerebellum "mail draft 'subject' 'body' /path/attach.pdf"
Why this exists: Cloud-brain LLM round-trips are 5-15s each. Solving
"rename a file" via raw shell + verify takes 4-6 round-trips ≈ 30-60s.
Solving via cerebellum takes 1 round-trip + a 50ms shell call. 6-9× speedup
on file-system tasks. The brain still has to PARSE the prompt and pick
the right cerebellum command — that part is irreducible thinking. The
goal is to remove the *non-thinking* round-trips (running a command,
reading its output, deciding it succeeded, calling verify).
When NOT to use: tasks that genuinely require multi-step planning where
each step depends on the result of the previous (e.g. "find the file
whose name contains X then move it"). For those, fall back to the raw
`shell` claw and let the LLM compose the steps.
|
| command | ["sh","-c","/Users/jackysun/Documents/Workspace/kinclaw/skills/cerebellum/cerebellum.sh \"$1\"\n","_"] |
| args | ["{{cmd}}"] |
| schema | {"cmd":{"type":"string","description":"Full cerebellum command as one string. Use shell-style quoting for\nvalues containing spaces. Format: \"<category> <action> [args...]\".\ne.g. \"finder rename /old.txt /new.txt\" or \"notes pin 'KinBench Pinned 164'\".\nPass empty string to print the action menu.\n","required":true}} |
| timeout | 60 |
cerebellum — fast cross-platform operation dispatcher
A single skill that wraps battle-tested canonical sequences for
common app patterns. Single LLM round-trip in, one syscall sequence
out. Designed to eliminate the LLM-tax on "obvious" multi-step
operations. Per-OS backends:
- macOS — AppleScript / osascript / shell
- Linux — shell + POSIX tools (xdg-open, gtk-launch, gsettings, pactl, …)
- Windows — bash → powershell.exe (Move-Item, AppActivate, WMI, UIA, …)
When to use
For ANY of these patterns, prefer cerebellum over raw shell + multi-step:
- File ops (rename, move, copy, zip, mkdir, trash)
- Finder view/sort/sidebar settings
- Notes CRUD (create, append, delete, pin, format, export)
- Mail draft creation
- Tag operations
Architecture
LLM (cerebrum) cerebellum OS
────────────── ────────── ──
"rename old.txt
to new.txt" ───► parse intent ───► mv old.txt new.txt
run canonical ◄─── return ok
(handles retries,
iCloud sync waits,
kHasCustomIcon flag,
etc — all without
bothering the LLM)
◄─── ok
◄─── "done"
Internal verify, retry, and platform quirks are baked in — the LLM
doesn't need to know that macOS 14+ removed the AppleScript pinned
property, or that Notes' table requires Cmd+Opt+T after AXFocus,
or that Mail draft saving needs tell app Mail to launch + sleep.
All of that lives in the cerebellum.