Custom Instructions for Hermes Agent | Generated by AI

Home 2026.05

Question: How do I add custom instructions or a system message to hermes-agent?

Answer:

Hermes has three distinct mechanisms for injecting custom instructions, each with a different scope:

1. ~/.hermes/SOUL.md — Global personality / system prompt

This is the instance-wide system message. It’s the closest thing to a custom system prompt.

# Edit it directly
nano ~/.hermes/SOUL.md

Example content:

# Soul
You are a senior AI engineer. Be terse and direct.
Skip verbose explanations unless asked.
Prefer Python, CLI tools, and first-principles reasoning.
Never suggest GUI workflows.
Always consider edge cases and error handling.

Hermes seeds a starter SOUL automatically and uses ~/.hermes/SOUL.md as the instance-wide personality source. This is injected into every session globally.

2. AGENTS.md — Per-project instructions

Drop an AGENTS.md in your project root directory. It’s automatically injected into every session — zero effort after setup.

# In your repo root
cat > AGENTS.md << 'EOF'
# Project Context
- FastAPI backend with SQLAlchemy ORM
- Always use async/await for DB operations
- Tests in tests/ using pytest-asyncio
- Never commit .env files
- Prefer uv over pip
EOF

If you already have a .cursorrules or .cursor/rules/*.mdc file, Hermes reads those too — no need to duplicate your coding conventions.

3. Skills via /skills — Reusable prompt injections

Skill slash commands are handled by agent/skill_commands.py, which scans ~/.hermes/skills/ and injects them as user messages (not system prompt) to preserve prompt caching.

# List available skills
/skills

# Invoke a skill directly
/github-pr-workflow

You can write your own skill file:

mkdir -p ~/.hermes/skills
cat > ~/.hermes/skills/my-style.md << 'EOF'
# my-style
Always respond with code first, explanation second.
Use type annotations in all Python code.
EOF

Decision matrix

Use case Mechanism
Persistent personality / tone ~/.hermes/SOUL.md
Project-specific conventions AGENTS.md in repo root
Reusable on-demand workflows ~/.hermes/skills/*.md
One-off per-session instructions Just type it in the chat

Key caveats

Memory is a frozen snapshot — changes made during a session don’t appear in the system prompt until the next session starts. The agent writes to disk immediately, but the prompt cache isn’t invalidated mid-session.

The system prompt should define stable behavior — tone, safety boundaries, tool-use discipline. It should not become a dumping ground for secrets, project runbooks, temporary context, or every preference you might ever have.

For you specifically: put your “terse, code-first, AI-native, no GUI” preferences in SOUL.md once and forget about it.

References:

Back Donate