Write Better Briefs and Control Cost with Hermes Commands

Last tested and updated: August 25, 2026, with Hermes Agent v0.20.5. Run /version first.

Four quiet commands improve everyday work from opposite ends: /prompt helps you give a better brief, /usage shows what this session consumed, /insights reveals longer-term patterns, and /model --once lets you spend extra capability on one hard turn instead of upgrading the entire conversation.

/prompt: write the brief in a real editor

The CLI-only /prompt command opens a temporary Markdown file using $VISUAL, then $EDITOR, with platform fallbacks. Seed it with a starting line if useful:

/prompt Draft a weekly research brief

Write something you can review before sending:

# Task

Compare three options for my weekly research workflow.

## Deliverable

- One-page summary
- A dated source link beside every factual claim
- Separate facts from judgment
- End with three recommended actions

## Boundaries

Do not sign into accounts, contact anyone, or use personal data without asking.

Save and quit; Hermes sends the body as the next turn. Lines beginning with #! are editor instructions and are stripped. Leave the body empty to cancel. The temporary file is deleted afterward, but submitted content still goes to the configured model provider.

/usage: check this session

/usage

This is a local/provider-account query rather than an LLM turn. Output varies by provider and surface, so do not expect every installation to show an exact dollar figure. Use it to notice a conversation that is growing expensive or context-heavy.

/insights: inspect patterns over time

/insights 7
/insights 30 --source cli

Insights reads Hermes’ local session database without an LLM call. It groups activity by models, platforms, tools, skills, and cost buckets. The value is not merely a total; it helps you spot jobs that use an expensive model repeatedly or tools that are barely used.

/model --once: escalate one turn

Use a configured model only for the next turn:

/model claude-opus-4-1 --once
Review this contract for contradictions and missed edge cases. Do not rewrite it yet.

After that turn succeeds, errors, or is interrupted, Hermes restores the previous model. You cannot combine --once with a persistent global switch. Switching breaks the prompt-cache prefix both when moving to the temporary model and when restoring it, so use it for genuinely hard review—not every routine formatting task.

A practical weekly routine

1. Use /prompt to write a source, deliverable, and boundary section.
2. Do routine collection with your normal model.
3. Use /model … --once for a high-stakes critique or synthesis.
4. Run /usage at the end of the session.
5. Review /insights 30 monthly and simplify repeated expensive work into skills or scripts.

The stable slash-command reference and command source define these behaviors.

Frequently asked questions

Can messaging users use /prompt?

No. It is a CLI-only editor workflow.

Which editor opens?

$VISUAL, then $EDITOR, then Notepad on Windows or Nano elsewhere.

Is Markdown required?

No, but headings and lists make a long instruction easier to review.

Does /usage use the model?

No. It queries local/provider usage information; fields vary by provider.

Does /insights upload my sessions for analysis?

The command reads the local SessionDB and does not make an LLM call.

Does --once permanently change the model?

No. Hermes restores the prior model after that single turn.

Why might the one-turn switch cost more than expected?

It uses the selected model and breaks prompt caching on switch and restore.