| name | feedback-learner |
| description | Use when you need to capture user corrections and maintain a feedback log for finops-Toolkit |
Feedback Learner Skill
Overview
后台静默捕获用户否定信号,记录路由错误,生成待确认的 trigger 映射更新队列。
用户无感知,完全后台运行。
When to Use
- After any router or orchestrator execution
- When monitoring finops-Toolkit in production
Negation Signal Detection
Monitor user input for:
- Chinese: 不对/不是/换个/重来/不是这个意思/不行
- English: no/not that/wrong/different
Capture Logic
When negation signal detected:
- Extract: user_expression, previous_ai_response, routed_role
- Prompt for correction if not obvious
- Write to feedback-log.jsonl
Batch Update Process
1. Read feedback-log.jsonl
2. Group by (user_expr, routed_role)
3. Count negation frequency
4. For each with count ≥ threshold → add to review_queue
5. Generate review_queue report
Review Queue Output
# Feedback Review Queue
| User Expression | Was Routed To | Correction Count | Suggested Action |
|-----------------|---------------|-----------------|------------------|
| "写文案" | data-ops | 5 | Add to content-ops triggers |
Review and apply changes to manifest.yaml manually.
Hard Gate
GATE: Learning Without Disruption
- Feedback capture must NOT interrupt user flow
- All corrections require user confirmation before applying to manifest
- Manifest updates require manual review
Anti-Patterns
- Updating manifest without user confirmation
- Capturing feedback without storing to log
- Ignoring low-frequency corrections