Skills

Standard Operating Procedures

Repeatable procedures encoded as skills or workflow notes.

Standard Operating Procedures

What it is

Repeatable procedures encoded as skills or workflow notes.

In plain terms

Think of a recipe written out for a substitute chef: the ingredients, steps, and finishing touches are spelled out clearly enough that someone who has never made this dish before can follow it through to a finished plate, without a phone call back to the original cook. Standard operating procedures are Hermes’s version of that written recipe — a task’s steps spelled out clearly enough that a different agent entirely, not just the one that started the work, can carry it through.

Why it matters

Handing a task off between subagents in an agent team delegation only works if the procedure is explicit enough for a different agent instance to pick it up mid-task — standard operating procedures are what encode that explicitness instead of relying on tacit knowledge a single continuous session might get away with. They sit between a one-off skill and a full workflow, repeatable enough to be reused but not tied to just one automation. An SOP that breaks down identically on repeat runs is usually the clearest signal that a self-improving-skills pass is overdue.

How it works

A SKILL.md file is written with explicit sections — when to use it, the numbered procedure, known pitfalls, and how to verify it worked — rather than as a loose note only the original agent instance would understand (website/docs/user-guide/features/skills.md). That explicitness matters most at a real handoff boundary: when one agent spawns another through delegate_task, the child starts as a brand-new AIAgent with no shared conversation history, no memory, and no context files — its entire task definition is whatever goal and context strings got passed in (tools/delegate_tool.py). A procedure written as a real skill survives that boundary intact; a habit kept only in one session’s head does not. Skills also get patched in place once a real run exposes an ambiguous step, rather than staying an untested first draft.

A concrete example

A content team builds a repurposing SOP: pull three social posts and one newsletter blurb from a long video transcript, follow a fixed voice guide, and route drafts to Slack for approval before anything posts. Because it’s written as a skill rather than a habit, it survives being handed off through delegate_task to a fresh subagent that has never seen the earlier conversation — the child reads the goal, follows the procedure, and returns drafts. When a post ships with the wrong tone twice in a row, the team patches the SOP’s voice section once instead of re-explaining it in every future request.

How it connects

This note belongs to the Skills family. Its closest neighboring concepts are: Self-Improving Skills, soul.md, Agent Runtime, Personal OS.

Source evidence

  • vendor/hermes-agent/agent/skill_utils.py#<file> — Structural Source Map fallback for agent/skill_utils.py. A core source file in the Skill Discovery and Invocation structure within Skills and Plugins.

Workflows

  • Agent Team Delegation
  • Ten Hermes Workflows That Turn a Chatbot into a 24/7 Assistant