Three Rules for Agent Skills, From the People Writing Them
Read the scripts, keep only what worked, and prune what the model now does alone.
Practical advice for writing and installing skills, one day after the standard moved to a foundation.
A skill is just a folder: a SKILL.md file plus scripts your agent may run, @CJF077 writes after Anthropic handed Agent Skills to the Agentic AI Foundation. That is why the roadmap item he cares about most is security guidance. “Until then, read the scripts before you install one.”
Anthropic’s announcement lists that roadmap: versioned releases, a formal proposal process, and security guidance for people who write and install skills.
On what to keep, @sudoingX is strict: “a skill should only stick after it worked on a real task, and the model never gets to grade.”
On what to drop, @rlaope tells Hermes Agent users to clean up their skills periodically. His first two criteria: skills that the raw model can now handle on its own, and skills made redundant by newly added core features.
| Rule | From |
|---|---|
| Read a skill’s scripts before installing it | @CJF077 |
| Keep a skill only after it worked on a real task, and do not let the model grade it | @sudoingX |
| Prune skills the model or the harness now covers | @rlaope |
Sources
- @CJF077 on X · Oct 22, 5:05 PM PT
- @sudoingX on X · Oct 22, 5:35 PM PT
- @rlaope on X · Oct 23, 12:38 AM PT
- Anthropic · anthropic.com