Update to Ciarem AI

{
  "description": "Use when an admin wants to set up, fix or improve their Ciarem AI agent, or says it answers something wrong, too long, too cold, or transfers when it shouldn't.",
  "included_files": [],
  "name": "tune-ai-agent",
  "skill_md_contents": "---\nname: tune-ai-agent\ndescription: Use when an admin wants to set up, fix or improve their Ciarem AI agent, or says it answers something wrong, too long, too cold, or transfers when it shouldn't.\n---\n\n# Tune the AI agent\n\nAdmins only: if a tool answers `requires_admin`, tell the user only workspace admins can change the AI agent.\n\n1. **Find what to fix.** Call `get_ai_setup`. If the user named no problem, also call `get_account_health`, `get_knowledge_gaps` and `get_ai_performance`, and propose the top three fixes. If they named a real conversation, call `explain_ai_decisions` for it.\n2. **Reproduce it.** Call `test_ai_reply` with a customer message that shows the problem, and show the reply before changing anything.\n3. **Pick the right lever**, one per problem:\n   - Tone, language, length, initiative, welcome: `update_ai_settings`.\n   - Who the business is and what it sells: `update_business_profile`.\n   - A missing or wrong fact (price, hours, policy): `add_knowledge_text` into a knowledge base the answering agent uses (`update_agent` `knowledge_base_ids`), or `delete_knowledge_source` for a wrong one.\n   - How to behave in a situation: `add_coaching`, with the customer message and the AI reply from step 2.\n   - When to hand over to a person: `create_transfer_rule` / `update_transfer_rule`.\n   - Which agent takes which conversations: `update_agent` `when_to_use`.\n4. **Apply it.** Say what you will change in one line, then call the tool. Change only what the user asked for.\n5. **Test it again.** Call `test_ai_reply` with the same message (a new `test_conversation_id` unless the problem needs the earlier turns), and compare before and after. If it isn't right yet, go back to step 3 with a different lever rather than piling on more corrections.\n6. **Check for regressions** after a round of changes: `run_ai_checks`, wait a few minutes, then `get_ai_check_results` and report what didn't pass.\n7. Close any recommended action the user fixed with `close_recommendation` (`resolve`).\n\nNever turn the AI off (`set_ai_active`) or delete anything unless the user asked for it.\n"
}