Feedback fix the tool not the prompt

When a bot's LLM makes a bad judgment call (e.g. giving up after one narrow search, or misformatting a query), Justin's default preference is to fix it deterministically in the underlying tool, not by appending another rule to the system prompt.

Why: Justin said explicitly, "seems like it should be smarter, don't want a shit ton of rules for every edge case." Prompt rules are per-case, don't generalize, and bloat over time; a tool-level guard (e.g. auto-broadening a Gmail search on 0 results before reporting empty) fixes the whole class of failure and benefits every persona that calls the tool.

How to apply: when diagnosing an LLM/bot misbehavior, first ask whether the fix belongs in the tool (deterministic, generalizes, no prompt bloat) before proposing prompt/rule additions. Only add prompt rules when the fix genuinely cannot be pushed down into the tool layer. This complements [[feedback_robust_over_quick]] and [[feedback_skill_systems_pattern]].

[auto-memory session 6b59fd6c-e0e6-4c96-a90d-d0dfe236981b, confidence 0.85, mode direct]