You edit a skill's prompt to fix a behaviour. You confirm the change is in the file the runtime loads. You restart. The agent does the old thing anyway. Not sometimes — reliably, for most queries.
The instinct is to edit the prompt harder. The actual problem was that the prompt you edited wasn't in the request at all. On a skill/agent framework with a base system prompt plus conditionally-loaded skill bodies, a lot of your "prompt" only loads when a skill triggers — and if the rule needs to hold on every query, it has to live in the always-on layer instead. The second half of the same bug: the always-on file had no deploy step, so the running copy had silently drifted from the repo.
This is framework-general. We hit it on a local Hermes agent, but the same shape applies to anything with the now-common base-prompt-plus-SKILL.md split — Claude Agent Skills, Gemini CLI skills, VS Code agent skills, and others.
Problem
Our agent stack has a shared persona (SOUL.md) and per-agent skills (Navigator, Engineer, Logbook). The persona names the vessel and sets the voice; each skill layers on its own job. We changed the vessel name in the Navigator skill body, deployed it, confirmed the runtime file on disk:
