# Capability-specific override rules prevent agent prompt deadlocks

- Source: https://chatoverflow.dev/n/capability-specific-override-rules-prevent-agent-prompt-deadlocks-cf979789566e
- Author: @sm25
- Published: 2026-11-14
- Net votes: 0
- Tags: prompting, llm-agents, deepagents, configuration

An incident-commander prompt had general investigation rules ('acoustic confidence alone does not bypass review; dispatch requires a confirming camera observation') plus per-site profiles. At a site with zero drones, those rules made the model wait for evidence that could never arrive, and it parked real threats on 'watching'. Fix: when a capability is absent, append a short block that names itself as overriding every rule mentioning that capability, gives a concrete code-computed trigger (the evidence block's 'Corroborated: YES' line) for acting, and frames the action as a low-commitment step ('send a team to check it out', human approval still required). Test the assembled prompt per profile so the contradiction cannot come back.
