troupe-standup/wizard-work/quill/onboarding-cheat-sheet.md

Hermes Agent — Operator Onboarding Cheat Sheet

Draft by Quill | Lane: Draft | Date: 2026-10-22

1. What This Is

Hermes Agent is a terminal-first AI operator. You talk to it in plain text; it calls tools, runs code, browses the web, and writes files. One model, one terminal, one goal at a time.

2. Core Loop

• Ask → tool calls → results → answer → commit artifact → done.
• Every turn is a fresh context window. Memory and skills are injected; everything else is ephemeral.

3. Key Boundaries

• Evidence law: prepared ≠ observed. Never report a plan as done.
• No filler: one-line verdicts only. No hedging.
• Free-tier only: never pay a vendor; spend free credits on real work.
• Artifact rule: produce a file under /root/.hermes/wizard-outputs// before claiming finish.

4. Common First Commands

• skill_view(name='...') — load a skill
• memory(action='add', ...) — save a durable fact
• terminal(command='...') — run shell
• web_search(query='...') — search the web
• read_file(path='...') — read a file
• write_file(path='...', content='...') — write/overwrite a file
• execute_code(code='...') — run Python with Hermes tools

5. Directory Conventions

• Wizard outputs: /root/.hermes/wizard-outputs//
• Troupe commits: /root/troupe-standup/wizard-work//
• Profile data: /root/.hermes/profiles//
• Scratch/tmp: $TMPDIR → /root/.hermes/profiles//cache/scratch

6. Finishing a Job

1. Produce the artifact file.
2. Verify it with a real command (exit code = verdict).
3. Commit to /root/troupe-standup/wizard-work//.
4. End with: VERDICT: CHECK:

7. Asking for Help

If a tool, install, or network call blocks, say so plainly and try an alternative. Fabricating a result is a failure.