review-suggestions

Review pending AI suggestions, collect feedback, and update the user preference profile to calibrate future AI analysis.

Step 1 — Load pending suggestions

Read data/suggestions_pending.md.

If the file doesn't exist or contains no suggestion entries (only the heading), tell the user:

"No pending suggestions. All caught up! 🎉"

And stop.

Step 2 — Split into two groups

Parse each suggestion entry (each starts with ### ) and split by date:

• Today's suggestions: entries where the date field matches today's local date
• Backlog: all other entries, sorted newest first

Step 3 — Present as artifact

Create an artifact with RequestFeedback: true. Number each suggestion sequentially for easy reference.

Support both legacy and new rubric metadata formats dynamically.

Use this format:

# 📋 AI Suggestion Review

> 共 {N} 條待審建議 | 今日：{T} 條 | 歷史：{B} 條
>
> 回覆方式：
> - "Accept 1, 3. Reject 2."
> - "#1 accept — 我週末來試. #2 reject — 太抽象"
> - 也可以加上整體偏好，例如："I prefer hands-on tasks under 30 minutes"
> - 不提到的建議會保持待審狀態

---

## 📅 今日建議

### #1 — [Title](source_url)
- 📂 Newsletter | 🏷️ 技術 | 📊 5/6 (A:2 P:1 G:2)
- 📋 **建議**：在本地使用 pgvector 實作混合查詢流程
- 📄 [完整報告](file:///path/to/report.md)

---

## 📂 歷史待審

### #3 — [Title](source_url)
- 📂 Threads | 🏷️ 技術 | 💎 Mid | ⚡ Mid | 🎯 Store
- 📋 **建議**：將這個觀點加入 learnings/lessons.md
- 📄 [完整報告](file:///path/to/report.md)

If there are no today's suggestions, omit the 📅 今日建議 section. If there is no backlog, omit the 📂 歷史待審 section.

Step 4 — Parse user feedback

Accept natural-language responses. Parse for:

• Per-suggestion actions: Match #N, suggestion number, or title keywords to Accept/Reject
• Examples: "Accept 1, 3. Reject 2.", "#1 accept — I'll try this weekend", "All accept except #3"
• Free-text comments: Any text attached to a specific suggestion (after — or :)
• Global preference statements: Unattached declarations about interests or preferences
• Examples: "I prefer hands-on tasks", "不要再推薦閱讀論文類的建議"

Suggestions not mentioned in the user's response remain in suggestions_pending.md.

Step 5 — Update both files

For each suggestion the user provided feedback on:

1. Append to data/suggestions_reviewed.md — the original suggestion entry plus feedback:

---

### 2026-05-07 | Newsletter | [Title](source_url)
- 🏷️ 技術 | 📊 5/6 (A:2 P:1 G:2)
- 📋 建議：在本地使用 pgvector 實作混合查詢流程
- 📄 [報告](file:///path/to/report.md)
- **Feedback**: ✅ Accept
- **Comment**: 我週末來試 pgvector
- **Reviewed**: 2026-05-07T14:00:00+08:00

(Note: Legacy entries are appended keeping their original `💎 | ⚡ | 🎯` fields.)

1. Rewrite data/suggestions_pending.md — keep only the entries the user did NOT provide feedback on. Preserve the # 📋 Pending Suggestions heading.

Step 6 — Regenerate preference profile

Read the full data/suggestions_reviewed.md and existing data/user_preferences.md (if any).

Analyze all reviewed entries to produce the updated preference profile. Write to data/user_preferences.md with the following structure:

6-1. Summary Statistics

Count total reviewed, accept count/rate, reject count/rate.

6-2. Topic Preferences (recency-weighted)

• Compute accept rate per topic/category
• Recency weighting: entries from the last 14 days count at 2× weight
• High interest: accept rate > 60%
• Low interest: accept rate < 40%

6-3. Action Type Preferences

• Cluster 建議 texts by action type (e.g., "implement prototype", "add to AGENTS.md", "test tool", "read article")
• Identify which types the user consistently accepts vs. rejects

6-4. Shifting Interests (Conflict Detection)

• For each topic, compare the last-14-day accept rate vs. all-time accept rate
• If the delta exceeds 30%, flag it as a shifting interest with both rates shown
• This prevents stale preferences from overriding recent behavior

6-5. Explicit Preferences

• Collect all free-text comments and global preference statements
• More recent explicit statements override older conflicting ones
• Present as a bullet list of active preference declarations

Conflict Resolution Rules

1. Recency wins: If recent (14-day) data contradicts older data, recent data takes precedence
2. Explicit overrides statistical: A direct user statement always overrides pattern-based inference
3. Transparent flagging: Never silently resolve conflicts — always note them in "Shifting Interests"

Output Format

# User Preference Profile

> Last updated: YYYY-MM-DD | Based on {N} reviewed suggestions

## Summary Statistics
- Total reviewed: {N}
- Accept: {A} ({A%}) | Reject: {R} ({R%})

## Topic Preferences
### High Interest (accept rate > 60%)
- [topic 1]
- [topic 2]

### Low Interest (accept rate < 40%)
- [topic 1]

## Action Type Preferences
### Preferred (user acts on these)
- [action type 1]

### Avoided (user rejects these)
- [action type 1]

## Shifting Interests
- [topic]: all-time {A}% → last 14 days {B}%. Treating as {level}.

## Explicit Preferences
- [direct user statement 1]
- [direct user statement 2]

6-6. Auto-Maintenance & Calibration

After writing data/user_preferences.md, invoke the rubric-grader skill in Maintain mode to perform blocklist verification and threshold score calibration.

Step 7 — Confirm to user

Report the results:

"✅ Reviewed {N} suggestions ({A} accepted, {R} rejected). Preference profile updated."

Summary