Skill Authoring Standards for AI Operators

Skills are Rocky’s procedural memory. They should capture reusable workflows that make future runs safer and faster, not temporary project status.

Official skills docs: https://hermes-agent.nousresearch.com/docs/reference/skills-catalog

Good skill content

• Clear trigger conditions.
• Numbered steps with exact commands when commands are stable.
• Public/private and safety boundaries.
• Known pitfalls from real runs.
• Verification steps that prove the workflow succeeded.

What not to save

• One-off task progress.
• Deployment IDs, PR numbers, issue numbers, or other soon-stale artifacts.
• Secrets, tokens, or credential locations.
• Raw customer data or private screenshots.
• Vague advice that does not change future behavior.

Authoring checklist

• Name the skill after the reusable workflow class.
• Keep project-specific lessons in references when the main skill would become too narrow.
• Patch outdated skills immediately when a real run reveals missing steps.
• Prefer concise declarative lessons over long transcripts.
• Include a verification checklist.

Pitfalls

• Creating a new skill for every completed task.
• Letting skills become stale and misleading.
• Mixing user preferences into procedural skills when memory is the better store.
• Saving private operational details in a skill that may load broadly.

Verification steps

• Re-read the skill as if a future agent is using it cold.
• Confirm it tells the agent when to use it and when not to.
• Confirm commands and paths are still accurate.
• Confirm it has pitfalls and verification, not just steps.