x-create

X Create

Create viral X posts (short tweets, threads, replies) based on user's persona and post patterns.

First-Time Setup

Check user profile before creating content:

Read references/user-profile.md
If initialized: false or file doesn't exist → Run onboarding
If initialized: true → Proceed to content creation

Onboarding Questions

Ask user these questions using AskUserQuestion tool:

Account Positioning (Niche): What does your X account mainly share?
- Options: AI/Tech, Entrepreneurship/Business, Personal Growth, Investing & Finance, Other

Target Audience: Who is your target reader?
- Options: Chinese-speaking users, English-speaking users, Bilingual users
Persona Style: What kind of persona do you want to build?
- Options: Professional and serious, Relaxed and humorous, Sharp opinions, Warm and approachable, Other

After collecting answers, update references/user-profile.md with initialized: true.

Post Types

5 Categories

Type	Style	Use When	Intent Signals (Routing Cues)
High-value substance	High information density, bookmarkable	Tutorials, tool recommendations, methodologies	Goal is bookmarks/retweets; emphasize actionable checklists, tools, steps
Sharp opinions	Has attitude and stance	Industry commentary, counterintuitive opinions	Goal is discussion/opposition; needs strong stance, contrast, counterintuitive
Trending commentary	Fast response	News commentary, event commentary	Goal is riding the trend/seizing timeliness; quickly interpret what just happened
Story insights	Personal experience + insights	Case studies, experience retrospectives	Goal is resonance/attention; use specific scenarios + twists + punchlines
Technical breakdown	Deep technical	Principle explanations, source code analysis	Goal is building expertise; explain principles, mechanisms, impact, and recommendations

Output Formats

Short tweets (≤280 characters) - Single tweet
Thread (multi-tweet thread) - 3-10 tweets connected
Comment replies - For replying to trending posts

Creation Workflow

Step 1: Load Context

1. Read references/user-profile.md → Get persona, style
2. (Optional) Read state from ~/.claude/skills/x-create/state/
   - liked_topics.json (positive samples)
   - rejected_topics.json (negative samples)
   - events.jsonl (optional)
3. Check assets/templates/{type}/ → Look for user reference posts
4. If no references → Use default patterns from references/post-patterns.md

Step 2: Intent-based Routing

Determine intent first, then choose style and format:

Intent → Style (5 categories)
- Bookmark/retweet oriented → High-value substance
- Discussion/opposition oriented → Sharp opinions
- Timeliness/trending oriented → Trending commentary
- Resonance/attention oriented → Story insights
- Professional/technical oriented → Technical breakdown
Style → Output format
- Short tweet: Single insight, quick take, one-liner
- Thread: Multi-point analysis, step-by-step, detailed breakdown
- Reply: Designed to respond to specific post/topic

If user explicitly provides --type, follow it. Otherwise route automatically.

Step 3: Apply Pattern

Read references/post-patterns.md for the specific post type pattern.

Step 4: Generate Content (A/B Variants)

Create two variants by default:

Variant A: More direct, stronger hook, higher contrast
Variant B: More structured, more evidence, slightly more neutral

Follow:

User's persona style
Selected post style pattern
Reference examples (if available)

Step 4.5: Humanize Pass (remove AI flavor, required by default)

For each variant, rewrite the text to sound like a real person on X while keeping meaning and claims unchanged:

Delete filler + chatbot politeness: avoid "of course/hope this helps/let's dive in"
Remove grand/marketing tone: avoid "marks/ crucial/ever-evolving landscape/demonstrates/empowers/breathtaking"
No vague attribution: avoid "experts believe/industry reports show" unless you provide a specific source; otherwise rewrite as "I've observed/my judgment is..."
Reduce connective phrases: avoid overusing "moreover/however/therefore"; prefer simple sentences and line breaks
Break formula: do not force "three-part structure"; 2 points is fine; mix short + long sentences
Avoid dash spam: do not stack "——"
Prefer concrete details over empty conclusions; if you are unsure, say it plainly and briefly

Thread constraints:

Each tweet must be <= 280 characters
Do not make every tweet identical in structure; allow 1-2 short "pause" lines

Step 5: Critic (Self-evaluation) + Rewrite Once

Score the humanized Variant A/B as the target reader (0-10):

Hook strength
Information density / value
Clarity and readability
Credibility (no exaggeration / no made-up facts)
Persona fit
Action likelihood: like / repost / bookmark / reply
"AI flavor" control: no empty grand statements, no templated endings, no vague authority

Rules:

If both Variant A and B score < 7, rewrite once (produce A2/B2), then run the same humanize pass again, and re-score.
Select the best variant as final output, but still show both drafts.

Output Format

# 推文创作

## 选题
{topic}

## 推文类型
{short_tweet/thread/reply}

## 风格
{post_style}

---

## Drafts

### Variant A

{For short tweet: single tweet content}

{For thread:}
### 1/N
{first tweet}

### 2/N
{second tweet}

...

### N/N
{final tweet with call to action}

**Critic score (0-10)**: {critic_score_a}

### Variant B

{For short tweet: single tweet content}

{For thread:}
### 1/N
{first tweet}

### 2/N
{second tweet}

...

### N/N
{final tweet with call to action}

**Critic score (0-10)**: {critic_score_b}

---

## Selected

Selected variant: {A|B|A2|B2}
Reason: {one-sentence reason}

---

## 发布建议
- 最佳发布时间: {suggestion}
- 配图建议: {image suggestion if applicable}
- 预期互动: {engagement prediction}

下一步：运行 /x-publish 发布到草稿箱

Append machine-readable blocks for hooks/state ingestion:

CREATE_JSON
{
  "schema_version": "x_skills.create.v1",
  "topic": "{topic}",
  "post_type": "short|thread|reply",
  "post_style": "high-value|sharp-opinion|trending-comment|story-insight|tech-analysis",
  "variants": [
    {"id":"A","critic_score_0_10":0,"text":"..."},
    {"id":"B","critic_score_0_10":0,"text":"..."}
  ],
  "selected": "A|B|A2|B2",
  "rewrite_once": true
}

HOOKS_JSON
{
  "schema_version": "x_skills.hooks.v1",
  "topic": "{topic}",
  "hooks": [
    {"text":"...","source":"variant.A","tags":["数字|反常识|痛点|悬念|类比"],"score_0_10":0}
  ]
}

Template Priority

User templates first: Check assets/templates/{type}/
Default patterns: Use references/post-patterns.md

Example:

Creating 高价值干货 post:
1. Check assets/templates/high-value/
2. If files exist → Learn style from examples
3. If empty → Use default pattern from post-patterns.md

Resources

references/user-profile.md

User customization info (shared across all x-skills)

references/post-patterns.md

Default viral post patterns for 5 categories

assets/templates/

User-provided reference posts organized by type:

high-value/ - High-value substance reference
sharp-opinion/ - Sharp opinions reference
trending-comment/ - Trending commentary reference
story-insight/ - Story insights reference
tech-analysis/ - Technical breakdown reference

Example

User: /x-create Claude 4.5 Opus发布 --type thread

Read user-profile.md → persona: Professional and serious, Sharp opinions
Check assets/templates/tech-analysis/ → empty
Read post-patterns.md → Get tech-analysis pattern
Generate thread:

### 1/5
Claude 4.5 Opus 上线了。我先说结论：它更像“慢一点，但更稳”的那类模型。

我用 3 个小任务试了下，写个线程记录👇

### 2/5
我最直观的感受不是“更聪明”，而是更会停下来检查自己。

同一个问题，它更少给“听起来对”的答案。

### 3/5
三个场景（都不算大项目）：
1) 重构一个旧模块：更愿意先问清边界，再动手改
2) 复杂推理题：会把关键假设写出来（这点很救命）
3) 长文档梳理：更少漏掉前后矛盾的地方

### 4/5
代价也很现实：
- 反应慢一点
- 成本可能更高（看你用的套餐/调用方式）
- 你得给它更明确的上下文

### 5/5
如果你做的是“错一次就很麻烦”的任务（代码、决策、长文整理），值得试。

只是日常闲聊，感知没那么强。你们试过了吗？

Integration

After creation, suggest:

推文创作完成！

- 类型: {thread/short/reply}
- 字数: {word_count}
- 预计阅读: {read_time}

下一步：运行 /x-publish 发布到X草稿箱

（反馈闭环，可选）
- 采纳并进入正样本：
  python ~/.claude/skills/x-create/scripts/x_state.py like --topic-json '{"title":"{topic}","selected":"{A|B}","critic_score":8}'
- 否决并进入负样本：
  python ~/.claude/skills/x-create/scripts/x_state.py reject --topic-json '{"title":"{topic}","reason":"low_value"}'
- 写入事件（hooks 自动收集也可用）：
  python ~/.claude/skills/x-create/scripts/x_state.py event --event create.generated --payload-json '{"topic":"{topic}","variants":["A","B"],"selected":"{A|B}"}'