Agent Skills - Procedural Knowledge for AI Agents

Skills are knowledge packages that teach Claude how to perform tasks better, not just tools that extend functionality.

Skills are bundles of “instructions, scripts, and resources” that Claude (or other agents via the open standard¹) to improve performance on specialized tasks.

Skills help to teach AI agents to complete tasks in a more repeatable fashion.

Skills follow the principle of progressivedisclosure

This concept has caught on so fast that the aforementioned Open Standard was created officially on December 18th, 2025.

”Writing” a Skill

The best way to get a feel of how skills work is to use Anthropic’s skill creator skill² to codify some procedural knowledge you already know, with the help of an AI coding agent which uses that skill.

In order to use the Skill Creator skill (using Claude Code as a harness for now) we actually need to install it (into the harness) since Claude Code by default doesn’t come with any skills. By default in Claude Code, skills are based on where they live.

In order to test this out using a global configuration of claude code, we can copy the skill creator skill to our global config as follows: cp -r /Users/svema/Repos/github.com/sushantvema/hackerspace/skills/skills/skill-creator ~/.claude/skills/skill-creator²

Then we can start a new claude code session (anywhere in your environment, but I’m preferring the ~/.claude directory for now), use the skills slash command, verify that skill-creator is enabled, and then ask Claude to create a simple skill.

Using the skill-creator skill or not, it is important to

Random

The agent skills reference³ contains a helper library called skills-ref which can be used to validate whether certain skills adhere to the specification. Although they claim that it should not be used in production, I have yet to try it out.

🌊 sushant's knowledge ocean

Recent Notes

PyO3

Agent Deck - Your AI Agent Command Center