Guides and Tips: Get More Out of Hermes in Daily Use

How To Actually Use Hermes Well#

A tool is only as good as the habits you build around it. The Hermes guides distil a set of practices that separate frustrated users from productive ones: clearer prompts, better project context, smart use of memory vs skills, token-budget discipline, and a few non-obvious operational security rules. This lesson pulls the most important ones into a single sitting.

Primary source: the tips page at https://hermes-agent.nousresearch.com/docs/guides/tips. Treat that as the authoritative source; this lesson is a summary and opinion layer.

Prompt Discipline: Specifics Beat Vagueness#

The biggest productivity win from day one is to front-load context. Instead of 'fix the code', try 'the auth/session.ts file throws TypeError: cannot read property id of null on line 47 when the session cookie is expired — handle that case by redirecting to login'. The more specific, the less the agent has to probe.

When you're debugging, paste the whole traceback. The agent parses it instantly; you save three rounds of 'what was the error?'.

`AGENTS.md`: Your Project's Institutional Memory#

Create an AGENTS.md at the root of your project. Put in it the things you'd re-explain every session — architectural decisions, naming conventions, testing approach, rules about which libraries to prefer. Hermes reads it automatically at session start. You stop repeating yourself, the agent stops proposing things you've already ruled out.

Keep it short and surgical. A bloated AGENTS.md is worse than no AGENTS.md — every line is in every prompt.

CLI Shortcuts That Save Real Time#

• Alt+Enter / Ctrl+J — newline without sending. Never accidentally ship a half-prompt again.
• Paste code directly — auto-detected and buffered as a single message, not fired line-by-line.
• Ctrl+C — interrupt the agent mid-operation and redirect instead of quitting.
• hermes -c — resume interrupted work with full history.
• Ctrl+V for clipboard images — paste a screenshot straight into the prompt.
• /title — name your sessions descriptively so hermes sessions list is useful later.

Memory vs Skills: Know the Difference#

Memory stores facts; skills encode procedures. When a multi-step workflow becomes repetitive — 'investigate an outage', 'run the release checklist', 'onboard a new engineer' — ask the agent to save it as a skill. Next time, the skill name gets you the whole procedure for free.

Memory has intentional capacity limits. When it fills, Hermes consolidates. You can also ask it to 'review and clean up memory' explicitly, which is a good monthly hygiene task.

Token Budget Control#

Token consumption matters in long sessions, and the prompt cache is the biggest lever. The cache rewards stable system prompts — switching model or context mid-session invalidates the cache and you pay full price. Decide on a model at the start, stick with it.

• /compress — summarise middle-of-session turns while keeping start and end intact.
• /usage — see where your tokens went before the budget runs out.
• delegate_task — split parallel investigations into subagents so only summaries return to the main thread.
• Ask the agent to write a unified Python script for bulk operations — cheaper than running twenty terminal commands one by one.

Operational Security#

A few rules to adopt immediately, before you're comfortable enough to get careless:

• Never set 'always' approval for destructive commands. Start with 'session' approval; upgrade deliberately.
• For untrusted repos, set TERMINAL_BACKEND=docker to contain the blast radius in a container.
• Never enable GATEWAY_ALLOW_ALL_USERS=true when the bot has terminal tools.
• Use per-platform allowlists (TELEGRAM_ALLOWED_USERS, etc.) or DM pairing codes for teammate access.
• Windows users: open files with encoding="utf-8" explicitly to avoid Unicode errors.

Personality: `SOUL.md`#

Drop a SOUL.md at ~/.hermes/SOUL.md to seed a consistent personality across every session — tone, verbosity, what to push back on. Combine it with per-project AGENTS.md for a layered 'global voice, local conventions' setup. Hermes even respects existing .cursorrules files if you already have one, so you're not duplicating.

Next Steps#

Lesson 7 is the developer-facing guide — architecture, how to contribute, how to extend. Lesson 8 is the CLI command reference you'll return to over time. If you'd rather skip the adoption curve altogether, SetupClaw's managed service applies all of the above — AGENTS.md conventions, security hardening, the layered-adoption principle — as part of the deployment.