A file that makes Hermes a careful coding assistant through SOUL.md: read first, small changes, run the tests, and never commit or delete without asking.
You use Hermes for code and want fewer surprise refactors, fewer claims that untested code works, and no risky commands without your say-so. Saves about 30 minutes
[LANGUAGES] | e.g. Python and TypeScript |
[EXPERIENCE_LEVEL] | e.g. intermediate, comfortable with git, new to Docker |
[CHECK_HABIT] | e.g. always run the test suite; if a project has none, run the changed script once |
[STYLE_PREFERENCES] | e.g. type hints everywhere, small functions, comments only where the code is not obvious |
Back these rules with settings: approvals.mode set to manual makes Hermes ask before every flagged command, and approvals.deny can block patterns such as force pushes outright. If you test risky code with terminal.backend set to docker, remember that container backends skip the dangerous-command checks, so the container is your only safety net.
Keep this SOUL.md general and put per-project details (run commands, folder map, off-limits files) in an AGENTS.md in each project, as in the AGENTS.md recipe in chapter 1.
Written for Hermes Agent v0.21.x. Check commands against your version with hermes --version. This site is independent and not affiliated with Nous Research.
Get the Hermes Starter Pack: 15 recipes for setup, SOUL.md, memory, skills, scheduled jobs, Telegram and approvals.
More free recipes: Write Your SOUL.md From Scratch · A SOUL.md for a Writing Partner
Need the whole set? The Hermes Agent Cookbook has 100 recipes for $27.