Free course · AI harness

Harnessing AI with Agent Skills

How to drive a coding agent with the site's Agent Skills — what a SKILL.md is, how one activates, and the guardrails, thresholds and evals that tell you the agent's work is right without reading every diff.

3 modules · 7 lessons · about 21 min of reading · no signup

Before you start: Finish one lesson of a layer course first, so the terms the skills encode are familiar. You need a terminal and any Agent-Skills-compatible coding agent.

Your progress 0 of 0 done (0%)

Start course

What you will be able to do

• Read and install a SKILL.md, and know which one an agent will load for the file it is editing.
• Write the guardrail before the skill, with thresholds a real codebase can pass.
• Catch an agent satisfying the measure instead of the goal, and score a skill with an eval set.

Course contents

Module 1 What a skill is

A procedure, a measurement and the loop between them.

1. Mark “A skill, a guardrail and a loop” complete 01 A skill, a guardrail and a loop Atomic Design · 3 min
2. Mark “Anatomy of a SKILL.md” complete 02 Anatomy of a SKILL.md Atom · 3 min
3. Mark “Install a skill and let paths activate it” complete 03 Install a skill and let paths activate it Reducer · 3 min

Module 2 Guardrails before skills

Decide how you would know the work is wrong, then make it executable.

1. Mark “Write the guardrail before the skill” complete 04 Write the guardrail before the skill Story · 3 min
2. Mark “Thresholds, baselines and scope” complete 05 Thresholds, baselines and scope Accessibility · 3 min

Module 3 Closing the loop

What an agent does to your checks, and how to score a skill instead of arguing about it.

1. Mark “When the agent satisfies the measure, not the goal” complete 06 When the agent satisfies the measure, not the goal Store · 3 min
2. Mark “Score a skill with an eval set” complete 07 Score a skill with an eval set Operations · 3 min