📈 Improves With Use — Muscle Memory
The central idea: the more you use Hermes and give continuous feedback, the better it gets. Like a muscle building memory, it learns your patterns, refines skills, and anticipates what you usually ask for. The gains are cumulative.
🔁 The improvement loop
- •Use → the agent executes
- •Give feedback → “this turned out well,” “I prefer it this way”
- •Iterates → next time, it works better
💡 Practical tip
Short, specific feedback is more valuable than generic praise. "Shorten the summaries to 3 bullets" builds muscle memory better than "it's great."
🎯 Ready-made skills and yours
Hermes has skills — saved routes for "how to do" a task. Some come ready-made and improve with use; others you build (see agentskills.io). A skill is a reusable recipe: it teaches the "how" once and you can reuse it whenever you need.
📊 Skill = the "how"
Think of a skill as a smart macro: you describe the procedure once, and the agent repeats it well. The more you use it, the sharper it gets.
🏛️ Pantheon — add personas
O Pantheon is where you create personas — dedicated specialists. Each persona (e.g., "the Alchemist") has job, description, system prompt and, crucially, a custom modelIt’s like putting together a team of specialists, each with the right brain.
Persona definition (illustrative)
persona: "o Alquimista" job: transformar ideias soltas em planos descrição: criativo, faz conexões inesperadas system_prompt: "Você é um sintetizador..." modelo: claude-opus (alto impacto)
💡 Practical tip
Create a few well-defined personas instead of many vague ones. A persona with a clear job + the right model delivers much more than ten generic personas.
🚗 Don't use Einstein to wash a car
The golden rule for saving money: use expensive models only for high-impact tasks. It doesn’t make sense to use a premium reasoning model (“Einstein”) for a trivial task (“wash the car”). On autopilot, use a cheap model; for what matters, use the strong model.
✓ The right model for the task
- ✓Difficult reasoning → Opus
- ✓Volume / routine → GPT / DeepSeek
- ✓Autopilot → low-cost model
✗ Waste
- ✗Opus for classifying email
- ✗Premium model in a routine loop
- ✗Same brain for everything
📊 Connects to Trail 1
That’s the multi-brain strategy: “if all you have is a hammer, everything looks like a nail.” Each persona/skill points to the model best suited for it—saving money without sacrificing quality where it matters.
🔬 "Deep research" skill — delegates to multiple models
Powerful example: the skill deep researchYou ask, "research the best country to live in," and it: starts a sub-agent, uses DeepSeek V4 + GPT in a loop for volume research, with Opus 4.7 reviewing the quality. Each model in the role it does best.
Starts a sub-agent
Fresh context dedicated to research, without cluttering the main conversation.
DeepSeek V4 + GPT in a loop
High-volume models collect and process data in iterations — cheap and fast.
Opus 4.7 reviews
The stronger model only comes in for the final review, ensuring quality where it matters.
💡 Why this is elegant
Combines skills + sub-agents + multi-brain. The expensive one (Opus) handles only what requires fine judgment; the cheaper one does the bulk of the work. Sub-agents are covered in Track 3.
⏹️ /stop — interrupt the task
Skills and personas give you a lot of power to act automatically — so you need a brake. The command /stop immediately stops the task in progress. Control is part of capability: being able to start isn't enough; you need to be able to stop.
/stop usage (illustrative)
› hermes: rodando skill deep-research (turno 7/20)… você: /stop › hermes: tarefa interrompida. 6 fontes coletadas até aqui.
📊 When to use /stop
- When you realize you asked for the wrong thing
- When a skill is using too many tokens
- When you want to redirect the agent before it finishes
📌 Module Summary
Next Module:
2.7 — 6 Power Keys: queue, background, compress