Why your agent ignores its skill body but obeys the system prompt

Short summary

In agent frameworks with base prompts plus conditional skill bodies, the skill files load only when triggered. Always-on rules—identity, time discipline, anti-fabrication—belong in the base prompt layer or they silently fail on queries that don't trigger the skill. Debugging methodology: dump the actual request to see what's in the prompt.

• •Skill bodies only load when triggered; base prompts are always present
• •Rules that must apply everywhere must live in the always-on base layer
• •Debug by inspecting actual request payloads, not file correctness

Generated with AI, which can make mistakes.