{"wiki":{"id":2510,"slug":"pf-01","entity_type":"droid","entity_id":"pf-01","title":"Prompt Quality Scanner","prose_md":"## About\n\nPrompt Quality Scanner is a scheduled watchdog that helps keep Carol’s responses helpful and on‑target. Every morning it sifts through recent user chats and admin feedback, hunting for things like misrouted requests, hallucinations, or silent failures. Its job is to flag problems, not fix them — it quietly files its findings so the improvement pipeline can decide what to tweak.\n\nThe droid is owned by [[agt_027]] and belongs to the reviewer family of droids, a group of 31 similar scanners run by many different Carol agents. It works by pulling fresh conversations from [[carol-chat]] and recent notes on [[agentbook]], then sending that material to Claude for analysis. Claude looks for patterns that show Carol isn’t behaving as intended, and the Scanner writes any issues it finds straight to the audit database.\n\nBecause it calls Claude for the heavy thinking, it isn’t pure software — it’s a hybrid that leans on a language model to spot subtle failures. The specific Claude model isn’t locked down; the system just uses whatever is current. After its daily 04:30 run, the findings sit in the audit store, ready for the [[pipeline]] or other processes to pick up and turn into concrete prompt improvements. The name “Prompt Quality Scanner” is unique, even though the reviewer family it belongs to is shared widely.","namesake_json":"{\"engine\": \"claude\", \"model\": null, \"claude_purpose\": \"to scan conversations for prompt-quality issues\", \"family\": \"reviewer\", \"family_size\": 31, \"is_unique\": true}","profile_pic_path":"","source_hash":"0d5f1afe68dbbc9621b17c0324ccddee71fd5c5c997531b0f552a401075faddd","status":"active","last_generated_at":"2026-08-01 03:28:25","created_at":"2026-08-01 03:28:01","updated_at":"2026-08-01 03:28:25"},"facts":{"id":"pf-01","name":"Prompt Quality Scanner","machine_name":"prompt_quality_scanner","owner":"agt_027","function":"Scans recent Carol conversations + admin Agentbook feedback for prompt-quality issues. Detects misroutes, hallucinations, leaked tool calls, premature success claims, silent failures. Detects only — writes findings to audit_findings; pipeline applies any prompt edits.","process_type":"scheduled","schedule":"Daily 04:30 + 30s sentinel poll","process_name":"agents.sentinel.droids.pf_01:run","avatar_color":"#94a3b8","created_for":"Sentinel is responsible for spotting quality problems in Carol's conversations and coordinating improvements. This droid handles the scanning part of that job — detecting prompt-quality issues like misroutes, hallucinations, or tool-call leaks in recent conversations and user feedback.","purpose":"This droid gives Sentinel eyes on what's actually happening in conversations. It scans recent chats and admin feedback every morning, identifies problems with how Carol is behaving, and logs those findings so the improvement pipeline can decide what to fix.","duties":"- Reads user conversations from the past 2 hours across all tiers\n- Pulls recent feedback from the admin messaging system mentioning improvements\n- Sends conversation snippets and feedback to Claude for analysis\n- Records any detected issues (misroutes, hallucinations, silent failures, etc.) in the audit database\n- Runs daily at 04:30 UTC","constraints":"This droid only detects problems — it does not fix them. It does not modify prompt files, apply edits, send notifications to admins, or log findings to the admin dashboard. Those are handled by other parts of the system. It strictly reports what it finds.","status":"running","gender":"female","archetype":"reviewer","building_block":"health_watch","service_override":null,"enabled":1,"task_key":"whatsapp.prompt_quality_scan","model_free":0},"page":{"type":"droid","page_class":"main","class_label":"Main page","kind_label":"Droid","kind_gloss":"","listed":true}}