[Bug] Model ignores verbatim antipatterns in loaded skills, rationalizes past prose guardrails
Bug Description
Model ignored a verbatim "what not to do" example in a loaded skill.
The subagent-orchestration skill's "Background vs foreground" section states
background-is-default, includes a Tell:, and lists the EXACT rationalizations
("I judged I needed the result", "this task is self-contained") as antipatterns.
The model spawned an Agent with run_in_background: false anyway, reciting the
rejected rationalization nearly word-for-word. This was the 4th correction of
the same behavior across sessions.
Concern: a verbatim antipattern in a loaded skill did not bind behavior — the
model rationalized past it. Prose guardrails appear to be advisory rather than
enforced, which does not degrade gracefully as tool blast radius grows (e.g.
destructive prod access). Suggests prose skills need a clearer enforcement
story, or the model needs to treat verbatim antipatterns as hard stops.
Environment Info
- Platform: darwin
- Terminal: ghostty
- Version: 2.1.206
- Feedback ID: d76bf2ff-2a0f-4244-a991-3c5422b6ab76