Finding
Hermes loses avoidable speed and cost gains when long prompts, cron prompts, and reusable task instructions change shape unnecessarily between runs.
Current
A real Hermes installation often uses repeated system instructions, profile prompts, cron jobs, skills, and long task templates. The weak point is usually prompt stability: useful instructions get rewritten ad hoc, examples are reordered, and recurring jobs include variable context too early in the prompt. That makes model-side prompt caching less effective and also makes agent behavior harder to compare across runs.
Suggested
- Keep stable instruction blocks stable. Exact change: add a “Prompt stability” rule to
SOUL.mdor the main profile instructions: “Do not rewrite stable role, safety, tool-use, and formatting instructions unless the behavior must change; put task-specific context after the stable reusable block.” - Standardize recurring cron prompt structure. Exact change: update the Optimizer Agent cron prompt template or cron runbook so every recurring job uses the same order: purpose, fixed scope, allowed tools, sources, change criteria, output format, then run-specific context.
- Move repeated long procedures into skills instead of pasting them into every task. Exact change: add a verification habit to the task completion runbook: “If the same long instruction block appears in two or more prompts, convert it into a skill or patch an existing skill, then reference that skill in future runs.”
Impact
This improves both operational consistency and model efficiency. Stable prompt structure makes repeated Hermes work easier for models to cache, which can reduce latency and cost on long or repeated calls when the provider supports caching. It also makes prompt changes easier to audit because meaningful edits stand out instead of being hidden inside routine rewrites.
Effort
Small — the work is mainly a prompt hygiene rule, a recurring-job template cleanup, and a habit of moving repeated instructions into skills.
Public page note
Safe public content includes the cache-aware operating principle, generic prompt-structure guidance, cron template discipline, and the recommendation to move repeated procedures into skills. Internal-only content includes exact private system prompts, raw cron prompts, private chat excerpts, provider billing details, logs, environment values, credentials, and any sensitive operational instructions.