Skops — AI skill for Claude Code

SkilloPS

**Loadout-based skill management for AI coding agents.**

Measure what every skill costs in context, activate only the skills a task needs, and keep the rest dormant.

Apache-2.0 · Node ≥22.6 · macOS / Linux / Windows

**Status: v0.3 — measurement works, and it corrected us.** `skops budget`, `skops lint` and `skops runtime` run today against a real skill install, in English or Chinese. `runtime` reads the listing Claude Code actually injected and found it **2.01× larger** than any disk scan can see. A read-only native macOS app is included. The activation mechanism loadouts depend on has since been **measured working** — see [Experiments](#experiments-the-activation-mechanism-works). See [Status](#status).

---

The problem

Every skill you install costs you twice.

**Once at rest.** A skill's `description` (plus `when_to_use`) is injected into the tool listing on *every single turn*, whether or not the skill is used. Anthropic's own docs put it plainly:

"Once a skill loads, its content stays in context across turns, so every line is a recurring token cost."

**Again in use.** The full `SKILL.md` loads on invocation and stays for the rest of the session.

Measured on a real machine

A disk scan of a real install reports this:

	Value
Always-on listing cost	1,623 tokens across 12 resident skills (~135 tokens/skill)
Total SKILL.md payload	178 KB across 12 files

But the disk scan is only half the story, and the missing half is the interesting one. Reading the listing Claude Code *actually injected* into a session on the same machine:

	Value
Skills in the injected listing	27
Skills visible on disk	20
Measured listing cost	3,257 tokens (13,025 chars)
Disk scan reported	1,623 tokens
Undercount	2.01×

**Fifteen of the 27 skills appear nowhere on the filesystem.** `dataviz`, `claude-api`, `update-config`, `code-review`, `run`, `loop`, `plugin-authorin