X Filter

Score and filter collected materials for X content creation. Topics scoring ≥7 points enter the creation pool.

Scoring System (满分10分)

Threshold: ≥7分 enters creation pool

Prerequisites

• Materials from x-collect (or manual input)
• User profile from x-create/references/user-profile.md (for relevance scoring)

Optional State (Feedback Loop)

If available, use persisted state to improve filtering:

• State dir: ~/.claude/skills/x-create/state/
• Negative samples: rejected_topics.json (SimilarityFilter)
• Events log: events.jsonl (optional analytics)

If state files don't exist, proceed normally.

Workflow

Input

Accept materials from:

1. x-collect output - Structured material report
2. Manual list - User-provided topics/URLs
3. Raw text - Unstructured content to evaluate

Scoring Process

For each material/topic, score multiple dimensions and combine with weights.

Weighted score (recommended):

FinalScore = Σ(w_i × s_i) - NegPenalty

Where:

• s_i are per-dimension scores (0..max)
• w_i come from references/user-profile.md (fallback to defaults)
• NegPenalty is derived from similarity to rejected topics and other negative signals

1. 热度/趋势 (Trending Score: 0-4)

4分: 当前热门话题，大量讨论
3分: 近期热点，关注度上升
2分: 稳定话题，持续有人讨论
1分: 小众话题，关注度有限
0分: 过时话题，几乎无人讨论

(Optional) 新鲜度 (Freshness)

• If you can infer recency from sources, either:
  • fold it into Trending score, or
  • add a small bonus/penalty (e.g., +0.5 for <72h, -0.5 for >30d)

2. 争议性 (Controversy Score: 0-2)

2分: 明显争议，多方观点对立
1分: 存在不同看法，可引发讨论
0分: 共识性话题，难以引发讨论

3. 高价值 (Value Score: 0-3)

3分: 硬核干货，可直接指导行动
2分: 有价值信息，提供新视角
1分: 一般信息，了解即可
0分: 低价值，无实质内容

4. 账号定位相关 (Relevance Score: 0-1)

1分: 与账号定位高度相关
0分: 与账号定位关联较弱

Check user profile at: ~/.claude/skills/x-create/references/user-profile.md If not found, assume domains: [AI/科技, 创业, 个人成长]

SimilarityFilter (Negative Samples):

• If ~/.claude/skills/x-create/state/rejected_topics.json exists, compare each candidate topic to rejected items.
• If max similarity ≥ 0.85: mark as likely duplicate/low-value → strong penalty or direct reject.
• If 0.75 ≤ similarity < 0.85: apply soft penalty (e.g., -2 points) and explain why.

(Implementation via script):

• python ~/.claude/skills/x-create/scripts/x_state.py similarity --against rejected --text "{topic}" --topk 3

Output Format

# 选题筛选报告

## 筛选时间
{timestamp}

## 用户定位
- 领域: {domains}
- 人设: {persona_style}

## 筛选结果

### Tier A：入选创作池 (≥7分)

#### 1. {Topic Title} - **{final_score}分**
| 热度 | 争议性 | 高价值 | 相关性 | 负向惩罚 |
|------|--------|--------|--------|----------|
| {trending}/4 | {controversy}/2 | {value}/3 | {relevance}/1 | -{neg_penalty} |

- **推荐类型**: [短推文/Thread/评论回复]
- **推荐风格**: [高价值干货/犀利观点/热点评论/故事洞察/技术解析]
- **创作角度**: 建议的切入点
- **核心观点**: 可提炼的关键论点
- **相似度命中（可选）**: {max_similarity} - matched: {matched_ids}

#### 2. ...

### Tier B：待定 (5-6分)
- {Topic} - {final_score}分 - {原因}

### Tier C：淘汰 (<5分)
- {Topic} - {final_score}分 - {原因}

## 创作建议

入选 {n} 个选题，建议优先级：
1. {最高分选题} - 理由
2. {次高分选题} - 理由

下一步：运行 `/x-create {选题}` 开始创作

Append a machine-readable block for hooks/state ingestion:

FILTER_JSON
{
  "schema_version": "x_skills.filter.v1",
  "timestamp": "{timestamp}",
  "profile": {
    "domains": ["..."],
    "persona_style": "..."
  },
  "items": [
    {
      "topic": "...",
      "scores": {
        "trending": 0,
        "controversy": 0,
        "value": 0,
        "relevance": 0,
        "neg_penalty": 0
      },
      "final_score": 0,
      "tier": "A|B|C",
      "reasons": ["..."],
      "similarity": {
        "max": 0.0,
        "matched": [{"id":"rej_xxx","score":0.0,"title":"..."}]
      }
    }
  ]
}

Execution Steps

1. Load materials from x-collect or user input
2. Read user profile for relevance scoring and weights
3. (Optional) SimilarityFilter against rejected topics
4. Score each material on criteria and compute FinalScore
5. Diversity adjustments (recommended):
   • Apply source/domain attenuation: score *= 0.6^(N-1) for repeated sources
   • Dedup per topic cluster: keep best-scoring item per cluster
6. Categorize: Tier A ≥7, Tier B 5-6, Tier C <5
7. Output report + FILTER_JSON
8. (Optional) Persist feedback-loop state:
   • Write Tier C items to rejected set:
     • python ~/.claude/skills/x-create/scripts/x_state.py reject --topic-json '{"title":"...","reason":"...","stage":"filter"}'
   • Append event:
     • python ~/.claude/skills/x-create/scripts/x_state.py event --event filter.scored --payload-json '{"accepted":3,"maybe":2,"rejected":7}'

Example

Input from x-collect:

素材1: Claude 4.5 Opus发布
素材2: AI编程助手对比评测
素材3: OpenAI最新裁员新闻

Scoring:

Claude 4.5 Opus发布:
- 热度: 4/4 (刚发布，热门话题)
- 争议性: 1/2 (性能vs价格讨论)
- 高价值: 3/3 (新能力详解)
- 相关性: 1/1 (AI/科技相关)
- 总分: 9/10 ✓ 入选

AI编程助手对比评测:
- 热度: 2/4 (持续话题)
- 争议性: 2/2 (Cursor vs Copilot争论)
- 高价值: 3/3 (实用对比)
- 相关性: 1/1 (科技相关)
- 总分: 8/10 ✓ 入选

OpenAI最新裁员新闻:
- 热度: 3/4 (近期热点)
- 争议性: 1/2 (有讨论)
- 高价值: 1/3 (信息价值有限)
- 相关性: 0/1 (非核心领域)
- 总分: 5/10 × 待定

Customization

Users can customize weights in user-profile.md:

scoring:
  trending: 4      # 热度权重
  controversy: 2   # 争议性权重
  value: 3         # 高价值权重
  relevance: 1     # 相关性权重
  threshold: 7     # 入选阈值

Integration

After filtering, suggest:

筛选完成！{n} 个选题入选创作池。

推荐优先创作：{top_topic}（{score}分）

下一步：运行 /x-create {top_topic} 开始创作