Appearance
Effective habits
Habits that make local Hermes genuinely useful day to day.
Core habits
- Verify offline setup first. 64k context + tool-calling model. Everything else builds on this.
- Start from project root. Context files load automatically.
- Brief like an agent. Concrete goals, not vague questions.
- Require approval for destructive ops.
approvals.mode: manual(orsmart) always. - Disable web toolsets when offline. Prevents wasted tool rounds.
- Iterate in session. Follow-ups beat restarting.
- Move repeats up the ladder. Prompt → context → skill → cron.
- Review agent-created skills. Before trusting them with shell access.
Habits by situation
Agent tries web search offline: Disable web toolsets in config.
Agent is slow: Use a smaller fallback (e.g. qwen3.5:9b) for simple tasks, Qwen 3.5 27B for agentic work. Schedule heavy jobs off-hours.
Agent forgets mid-task: Check Ollama context length. Likely set to 2048.
Agent only chats, no tools: Model lacks tool calling. Switch to qwen3.5:27b or similar.
Quality is poor: Try larger model, or add cloud fallback for hard tasks only.
Weekly practice
- Review
~/.hermes/memory/for stale entries. - Check agent-created skills for accuracy.
- Run one use case you haven't tried yet.
- Note what you repeated and promote it to a skill.
The one-sentence summary
Configure offline correctly once, then systematize everything you do twice.