Local only
On this device
0 / 13 proofs · 0%
Part II · Compounding harness · 09

Skills

On-demand procedural memory, not a second system-prompt dump.

20 min

What a skill is

Skills path
01
Catalog
Names only
02
Select
Right skill
03
Body loads
Procedure
04
Run
Proof file

A skill is a reusable instruction package (SKILL.md). Front matter names and describes it; the body teaches the procedure. Hermes sees names and descriptions first: a vague description means the right skill may never load.

Skills use progressive disclosure: the catalog stays cheap; only the selected body enters the prompt. That is how a large library avoids burning the context window every turn.

  • Skill = procedure / knowledge package (agentskills-compatible).
  • Plugin / MCP = executable surface or product hooks, a different layer.
  • Slash commands (for example /plan) and /learn can capture workflows as skills when supported.
  • Install narrowly; run security scans on hub skills before trust.
  • Pin or curator: critical skills should not vanish to automated cleanup when you outgrow defaults.

Desktop first

Steps · tap to check0/4
If every skill is enabled, the model spends tokens choosing rather than working. Least privilege. Skills Hub is a map, not homework.
Skills system ↗ Work with skills (guide) ↗ Curator (later) ↗

Skills tax the fixed budget

The skills index and every enabled tool schema load into the empty-session budget (see module 06, prompt-size). A large "install for later" catalog costs tokens before any task starts.

Steps · tap to check0/3
Lab: Prompt budget audit →

Proof

Checklist · this device0/4
Next: Security → Sample skills-notes ↘
Proof of completion

You listed skills for this profile, ran or inspected one, and wrote skills-notes.md with name and when to use it.

Proof · this device0/1

Stored on this device only, not sent anywhere.