A SOUL.md file gives an AI agent its personality: who it is, how it speaks, what it refuses to do. A good one is short, specific, and written for your work — not a wall of adjectives. Here is the structure that holds up in daily use, with a skeleton you can copy.
SOUL.md is a plain text file that agent tools read when your agent starts. It comes from the OpenClaw ecosystem (an independent open-source project — AfterHours is not affiliated with it) and the convention has spread: one file that holds the agent's identity, separate from its task instructions. The agent's tone, its limits, its manners with you — all of it lives here.
Every section answers one question the agent will face daily. Skip the poetry; an agent with a vague soul gives vague answers. The skeleton below is the minimum that works:
Paste this into a file named SOUL.md and replace every bracket. Each line is deliberately short — you'll expand what your daily use proves necessary, not what sounds nice on day one:
Most shared templates fail the same way: a generic personality produces a generic agent. "Helpful, friendly, proactive" describes every assistant ever shipped — so the agent behaves like none in particular. The fix is specificity: your clients, your tools, your red lines. A SOUL.md written for a freelance designer should be useless to an accountant. If yours could belong to anyone, it belongs to no one.
Writing a good identity file takes a few honest hours. AfterHours agent files ship with the complete equivalent built in — identity, memory, safety rules — and tune themselves to you through a few questions at install: you paste one file into Claude Code or Codex, the tool sets everything up, no terminal. Different route, same destination: an agent that sounds like it works for you, because it does.
A plain text file that defines an AI agent's personality: identity, voice, values, refusals, and how it behaves when unsure. Agent tools read it at startup, separately from task instructions.
SOUL.md says who the agent is; AGENTS.md says how it should work — tools, rules of the workspace, task conventions. Personality in one file, operations in the other. Mixing them is the most common template mistake.
Tools accept long files, but as a working rule the strong ones fit on one screen: identity in two sentences, voice with one example, a handful of hard refusals. Short files stay maintainable — and you can actually tell when one line stops being true.
They include the complete equivalent for their runtime: a finished identity, memory and safety rules, written in, and adjusted to you by a few questions during install. You don't write it — you answer questions and the file does the rest.