social-media-title-insight

Analyze social media post titles to discover what makes content perform well. Use when user uploads data (Excel, CSV, JSON, text) with titles and engagement...

Security Scan

The skill mostly matches its stated purpose (analyzing titles and producing reports), but the bundled scripts contact an external service with inconsistent endpoints/tenant IDs and have undocumented dependencies — these mismatches and network calls warrant caution before installing or using the skill with sensitive data.

Purpose & Capability

The name/description (title-insight from uploaded data or by pulling an account) aligns with the provided scripts: data parsing, feature extraction, verification, and HTML report generation. However there are notable inconsistencies: the SKILL.md documents an API endpoint at https://vms-service-tx.tezign.com with header x-tenant-id: tx_t1, while the data_tool.py source contains a different URL (https://vms-service.tezign.com) and default tenant_id='t221'. html_to_pdf.py requires Playwright (commented dependency) but the SKILL.md only instructs installing pandas/openpyxl. These mismatches are unexpected and unexplained.

Instruction Scope

SKILL.md instructs running the included scripts and describes an API-based 'preview' mode. The scripts will create run directories and a marker file (.social-media-title-insight-latest-run) in the working directory. The fetch_api function performs POST requests to an external service to fetch account data (sending account names). The skill does not request credentials, but the code accepts an optional cookie parameter and sets a tenant header — the documentation and code disagree on endpoint and tenant values. The instructions do not clearly state what data is sent to the remote API (beyond account names), nor do they explain whether any uploaded local files are transmitted remotely; this lack of clarity about network behavior is concerning for privacy-sensitive data.

Install Mechanism

No formal install spec is provided (instruction-only plus bundled scripts). SKILL.md tells users to pip install pandas and openpyxl, but html_to_pdf.py requires Playwright (not documented). The scripts will run from disk, so missing or undocumented dependencies (Playwright, possibly other libs) may cause silent failures or unexpected prompts to install additional packages. No remote code downloads are present in the install spec, but the tool relies on running local Python scripts that perform network requests.

Credentials

The skill declares no required environment variables or credentials, which is consistent with the published metadata. That said, the code uses an external API and accepts an optional cookie parameter and tenant header — while not required, these could be used to pass credentials to the upstream service. The lack of any declared credential requirement but presence of optional cookie/tenant handling in code is an inconsistency to be aware of.

Persistence & Privilege

The skill is not always-enabled and does not request elevated agent privileges. It writes run directories and a marker file in the current working directory to track the latest run; this is limited, local persistence and consistent with its purpose. There is no evidence it modifies other skills or global agent config.

What to consider before installing

This skill appears to implement the advertised title-analysis workflow, but exercise caution: (1) The bundled script will contact an external service (tezign.com) when you use the account-preview mode — confirm you are comfortable sending account names and any other data. (2) SKILL.md and the code disagree on which endpoint and tenant ID are used; ask the publisher which endpoint is correct and what data the API receives and returns. (3) The HTML-to-PDF script requires Playwright (not documented in SKILL.md); run in an isolated environment or container and inspect/ pin dependency installs before running. (4) If you will analyze sensitive or private content, prefer using the local-file workflow (preview/auto with local input) and avoid the API mode until you verify the remote service's privacy policy and data handling. (5) Consider opening the full data_tool.py (complete file) and reviewing exactly what is POSTed to remote endpoints, and run the script with network access disabled if you want to guarantee no outbound calls.

社媒内容标题洞察分析

从高赞内容中发现规律，对比低表现内容验证差异，输出专业 HTML 报告。

核心理念

LLM只做最轻的"发现特征"，统计和报告生成全部由脚本完成。

工具依赖

pip install pandas openpyxl --break-system-packages -q

脚本位置：scripts/data_tool.py（与本SKILL.md同目录）

---

执行流程（5步，每步极轻量）

全自动模式（推荐：上传文件/粘贴数据）

用户上传文件或直接粘贴数据时，可直接一键分析（自动识别列、自动生成特征、自动验证、自动出报告）：

# 方式1：上传文件
python {SKILL_DIR}/scripts/data_tool.py auto --input {文件路径} --name "账号名"

# 方式2：直接粘贴（JSON/CSV/TSV/逐行文本）
python {SKILL_DIR}/scripts/data_tool.py auto --paste "$(pbpaste)" --name "账号名"

输出将写入 ./runs/<timestamp>/：

• _data_cache.csv
• _features.auto.json
• _verify_result.json
• report.html
• auto_detect.json

Step 1: 预览数据

RUN_DIR=./runs/$(date +%Y%m%d-%H%M%S)
python {SKILL_DIR}/scripts/data_tool.py preview --accounts "账号名(平台)" --size 100 --run-dir "$RUN_DIR"
# 或
python {SKILL_DIR}/scripts/data_tool.py preview --input {文件路径} --run-dir "$RUN_DIR"

看输出，判断哪列是标题、哪列是指标。数据会自动缓存到运行目录的 "$RUN_DIR/_data_cache.csv"。
如果不传 --run-dir，脚本会自动创建 ./runs/<timestamp>/ 并默认沿用最近一次运行目录。

Step 2: 排序取样

按识别出的指标列排序，看Top和Bottom标题：

python {SKILL_DIR}/scripts/data_tool.py sort --input "$RUN_DIR/_data_cache.csv" --col "engagement" --title-col "title" --n 25

数据量<100 用 --n 2025，100500 用 --n 30~50

Step 3: 输出特征JSON（LLM唯一核心任务）

仔细对比Top和Bottom标题，发现差异。然后将发现写入JSON文件，格式如下：

cat > "$RUN_DIR/_features.json" << 'FEATURES_EOF'
[
  {
    "label": "实用指南型",
    "description": "包含具体的选购指导或穿搭教程",
    "match_keywords": ["如何", "怎么选", "指南", "攻略", "教程", "法则"]
  },
  {
    "label": "明星/IP联名",
    "description": "标题中含有明星名字或联名品牌",
    "match_keywords": ["付航", "联名", "携手", "×"]
  },
  {
    "label": "系列栏目化",
    "description": "北面硬壳｜ 这类固定格式的系列标题",
    "match_keywords": ["硬壳｜", "巅峰系列｜", "联名｜"]
  }
]
FEATURES_EOF

每个特征包含：

• label：3~6字特征名
• description：一句话说明
• match_keywords：用于在全量数据中匹配的关键词列表（标题包含其中任一即命中）

特征发现角度（不限于）：

• 用词选择、表达方式、句式、语气
• 长度、标点、emoji、数字
• 情绪调性、内容策略、选题角度
• 人称、格式、栏目化

目标：发现10~15个特征。

Step 4: 脚本做定量验证

python {SKILL_DIR}/scripts/data_tool.py verify \
  --input "$RUN_DIR/_data_cache.csv" \
  --features "$RUN_DIR/_features.json" \
  --output "$RUN_DIR/_verify_result.json" \
  --run-dir "$RUN_DIR"

脚本会自动对每个特征计算：含特征 vs 不含特征的互动量/转赞比差异，输出结构化JSON。

Step 4.5（可选）: 输出定性洞察JSON

看完verify摘要后，如果有因果辨析的洞察，写入JSON：

cat > "$RUN_DIR/_insights.json" << 'INSIGHTS_EOF'
[
  {
    "title": "热度≠传播力",
    "content": "互动引导型标题热度高但转赞比低，吸引的是参与型用户而非传播型用户",
    "importance": "high"
  },
  {
    "title": "系列栏目化的真实价值",
    "content": "北面硬壳｜系列热度中等但转赞比高，说明栏目化内容吸引的是高质量粉丝",
    "importance": "normal"
  }
]
INSIGHTS_EOF

Step 5: 脚本生成HTML报告

python {SKILL_DIR}/scripts/data_tool.py report \
  --verify-json "$RUN_DIR/_verify_result.json" \
  --name "TheNorthFace" \
  --output "$RUN_DIR/report.html" \
  --insights "$RUN_DIR/_insights.json" \
  --run-dir "$RUN_DIR"

然后将报告复制到输出目录供用户下载。

---

数据来源

API拉取（用户未上传文件时）

python {SKILL_DIR}/scripts/data_tool.py preview --accounts "账号名(平台)" --size 100

• 接口：POST https://vms-service-tx.tezign.com/datacenter/ai-insight/public/account-data?size=N
• Header：x-tenant-id: tx_t1，Content-Type: application/json
• Body：["TheNorthFace(小红书)"]
• 返回：title, hot(热度), rate(转赞比%), account, author

本地文件

支持 Excel/CSV/TSV/JSON/TXT：

python {SKILL_DIR}/scripts/data_tool.py preview --input {文件路径}

---

脚本命令速查

命令	用途	谁做
preview	查看数据结构	脚本
sort	排序取Top/Bottom	脚本
compute	多列加权排序	脚本
verify	定量验证特征	脚本
report	生成HTML报告	脚本
auto	全自动分析（识别+验证+报告）	脚本

LLM只做：看数据 → 写 _features.json → 可选写 _insights.json

---

多指标分析

如果数据有多个指标列（如热度+转赞比），可以：

1. 先按热度 sort 一轮，发现特征
2. 再按转赞比 sort 一轮，补充特征
3. 合并所有特征到一个 _features.json，一次 verify 即可（verify会同时计算两个指标的差异）

---

异常处理

• API失败 → 提示用户手动上传数据
• 数据量<20 → 提示统计意义有限，但仍分析
• verify没有显著特征 → 降低 --min-diff 10 重试
• 特征命中数不足 → 调整match_keywords更宽泛