Yes, changing the underlying LLM model in Hermes often changes how skills (and tools) are processed. This is a common experience for users like you building with Hermes/OpenClaw.

Why This Happens
Hermes is model-agnostic at the framework level — you can swap models via hermes model without rewriting code. However, the actual behavior of skills and tool-calling changes because:
Different models have varying tool-calling / function-calling quality — Stronger models (e.g., Claude variants, Qwen3.5/3.6, certain GLM) follow structured tool schemas more reliably, make fewer malformed calls, and chain tools/skills better. Weaker or smaller local models (e.g., some Gemma versions) hallucinate parameters, skip steps, or fail to invoke the right skill.
Reasoning and skill application differ — Hermes skills are reusable Markdown procedures (step-by-step workflows the agent learned). A high-capability model can interpret, adapt, and combine them more intelligently. A weaker model treats them more literally or misses nuances, leading to different execution paths.
Context handling & prompt interpretation — Models vary in how well they use the injected memory, skill index, and system prompts. Switching can make the agent "forget" how to apply a skill correctly until the session resets or it re-learns.
Session / caching effects — Model changes usually require a /reset or new session for full effect (to clear cached prompts/tools). Without it, behavior can be inconsistent.
Common Observations from Users
Switching to a strong tool-calling model (like Qwen or Claude) makes skills feel much more reliable and autonomous.
Dropping to a smaller local model often makes complex skill chains break or become slower/less creative.
The self-improvement loop (auto-creating/refining skills) also performs differently — better models generate higher-quality skills.
Quick Fixes / Best Practices
Use hermes model to switch, then /reset the session.
Test skills right after switching — ask Hermes to evaluate or re-run a recent task.
Pin good models for skill-heavy work (many users recommend specific Qwen or GLM variants for local + strong tool use).
You can even experiment with per-skill model routing in some setups, though it's not fully seamless yet.