菜单
Test if the user's confidence matches their actual knowledge. Use when calibration ECE is high or strategy is "overconfident".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | calibration-check |
| description | Test if the user's confidence matches their actual knowledge. Use when calibration ECE is high or strategy is "overconfident". |
| user-invocable | false |
Detect and correct the gap between what the user THINKS they know and what they ACTUALLY know.
curl -s -H "x-api-key: $TOKEN" \
"https://cognition-api.fly.dev/v1/calibration/snapshot?user_id=USER_ID"
Pick 5 concepts — mix of high and low retrievability
For each concept, ask TWO things:
Concept: [X]
First: How confident are you that you could explain [X] right now? (1-5)
Now: [actual question about X]
After all 5:
Calibration Results:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Concept Confidence Actual Gap
useEffect cleanup 5/5 ✗ OVERCONFIDENT
SQL JOINs 2/5 ✓ UNDERCONFIDENT
Git rebase 4/5 ✓ Calibrated
Python decorators 3/5 ✗ Calibrated
CSS Grid 4/5 ✗ OVERCONFIDENT
Overall: You tend to overestimate your recall.
ECE: 0.32 (needs work — below 0.15 is well-calibrated)
Coaching:
Record all results as quiz_answer events with confidence_self_report values.
Overconfidence is the #1 enemy of learning. If you think you know something, you won't review it. If you don't review it, it decays. The calibration check breaks this cycle.
Connect a notification channel (Slack, Gmail, etc.) so Cognition can send you learning exercises throughout the day — even when your laptop is off.
Schedule learning exercises for the next 24 hours via Slack/Gmail. Runs at the end of every session so exercises arrive even when Claude Code is closed. The API decides WHEN and WHAT technique — you build the exercise and schedule delivery.
Send screen observations to the Cognition API as learning events. Transforms Screenpipe data into API event format.
Launch the Cognition GUI app. Opens the Electron window and connects Claude Code to it via WebSocket. Use this after /cognition:start or whenever you need the GUI.
Generate and deliver a learning exercise based on what the user is forgetting. Orchestrates different learning techniques.
Quick surprise retention check. A 30-second mini-quiz to test if a concept is still in memory. Use for concepts with high stability that haven't been tested recently.