Finding
Claude-style reasoning is operationally weak when it is used as an occasional model preference instead of a defined specialist lane for high-reasoning work.
Current
A real Hermes installation may route difficult analysis, code review, LangGraph/LangChain design, or adversarial critique to a stronger reasoning model or specialist subagent when the operator remembers to do so. The weak point is consistency: without a trigger rule, the system can underuse deep reasoning on tasks that need it, overuse it on simple work, or fail to capture why the specialist model changed the outcome. That makes model routing feel like taste instead of operations.
Suggested
- Define the Claude-specialist trigger boundary. Exact change: add a “Model Lore routing” section to
SOUL.mdor the main profile instructions: “Use a Claude-family or equivalent high-reasoning specialist for deep code review, LangGraph/LangChain architecture, prompt-risk analysis, multi-step debugging, adversarial critique, and ambiguous design trade-offs; do not use it for simple lookup, formatting, or routine edits.” - Create a reusable Claude-style review prompt. Exact change: add a skill or prompt template named
claude-specialist-reviewwith required sections for task context, assumptions, architecture critique, failure modes, alternative design, safety/privacy concerns, and final recommendation. - Add a model-lane verification habit. Exact change: update the task completion runbook or Optimizer Agent review prompt with: “When a high-reasoning specialist lane is used, record the task class, why escalation was justified, what decision changed, and whether the output should become a skill, memory, or artifact.”
Impact
This makes advanced reasoning a controlled capability rather than an ad-hoc model switch. Hermes gets deeper review where it matters most, especially for architecture, code quality, and LangGraph/LangChain design, while avoiding unnecessary token spend on routine work. It also creates an evidence trail for whether specialist routing improves decisions enough to justify continued use.
Effort
Small — the change is one routing rule, one reusable specialist prompt or skill, and one lightweight verification habit. No new infrastructure is required if a Claude-family or equivalent high-reasoning model lane is already available.
Public page note
Safe public content includes the maturity principle, generic model-routing rules, prompt-template structure, review categories, and operational benefits. Internal-only content includes provider credentials, private model configuration, raw review outputs, proprietary code, private LangGraph/LangChain designs, customer data, logs, and exact internal routing history.