Skill Library Field Workshop¶
Run your agent skill library like an operator: measure every skill against the task it claims to help with, ship only on evidence, and retire the ones that stop earning their context cost.
What makes this different
Four ideas run through every module and exercise. This is where the "write your first skill" tutorials stop and this workshop starts.
Ablation as a ship gate¶
A skill's value is the delta between the agent with it and without it. You run skills with-vs-without and read the lift before anything is allowed to ship.
Eval-as-contract¶
EVAL.yaml isn't a test you write after the fact — it's the contract for what the skill is for. A failing eval is a spec violation, not a flaky test.
Governance tiers¶
Experimental, team, and org-default skills — each with ownership, a review policy, and a promotion path between tiers. Not every skill deserves the same trust.
The retirement stage¶
Libraries rot. You practice the full create → review → retire lifecycle, including the call to kill or consolidate a skill that no longer earns its keep.
The six modules
Concept first, then a hands-on exercise for each. Work them in order.
Module 1 — Why Skills
A skill's value is the delta it creates. Learn to tell a skill that earns its context cost from one that quietly makes the agent worse.
Module 2 — Anatomy of a Skill
The three progressive-disclosure layers, and how to match each piece of content to the cheapest layer that can hold it.
Module 3 — Writing Skills That Trigger
The description is a trigger, not a summary. Signal versus noise, and why reaching for caps-lock is a smell.
Module 4 — Designing the Library
Three governance tiers, config inheritance, and the context economics of curating an org-default set.
Module 5 — Evals and Ablation
The eval is the skill's contract. Task tests versus quizzes, and ablation as the ship gate that proves lift.
Module 6 — Governance and the Lifecycle
Ownership, review policy, and the create → review → retire lifecycle that most guidance skips entirely.
Before you start
Ten minutes of setup: clone the repo, create a Python environment, and confirm the three gates (sanitize, validate, eval) run green. Begin on the Setup page.
Practitioners who already have agents in the loop and are accumulating skills faster than they can reason about them. No specific vendor's agent required — everything runs in demo mode on the open Agent Skills standard.
Not a "write your first SKILL.md" tutorial and not a skill directory or marketplace. We start where those stop and teach the operating discipline that keeps a library healthy.