Ai Daily

Generate daily AI news digest from feedship subscriptions. Use when user wants today's news summary, daily briefing, periodic news recap, AI daily digest, AI...

Security Scan

The skill's behavior mostly matches its stated purpose, but several instructions reference user-specific paths and OS-specific commands and there are a few sloppy assumptions that could lead to misuse or surprises.

Purpose & Capability

The skill claims to generate a daily AI digest from Feedship subscriptions and its scripts/instructions are consistent with that goal (calls feedship, filters articles, sends prompts to an LLM, and runs a local replacement script). Requiring the 'uv' binary aligns with the 'uv run feedship' usage. However, SKILL.md contains hard-coded, user-specific paths (e.g., /Users/y3/feedship) and assumes a local feedship checkout; these are not explained or parameterized and therefore surprising for a reusable skill.

Instruction Scope

Runtime instructions instruct the agent to cd into a specific user path (/Users/y3/feedship), run local feedship commands, write/read files under /tmp, and call an LLM with a local prompt template. The hard-coded path and macOS-specific date flags (date -v-1d) are environment assumptions that may fail or cause the agent to access unintended local data. The cron example posts outputs to channels (--channel / --to), which is expected for a reporting skill but should be noted as outbound publication points controlled by the user.

Install Mechanism

No install spec is included (instruction-only plus a small local script). This is low-risk: nothing is downloaded or executed from remote URLs and the included replace_refs.py is small and readable.

Credentials

The skill declares no required environment variables or credentials, and the only env usage (TODAY_ARTICLES) is set inline when invoking the replacement script. That is proportionate. Still, the skill implicitly depends on the user having a local feedship repo and valid feedship commands available; those are not surfaced as explicit requirements besides the 'uv' binary.

Persistence & Privilege

always is false and the skill does not request persistent/always-on privileges. It does show how to register a cron job via openclaw (user-controlled), but it does not modify other skills or system-wide settings beyond the usual cron registration.

What to consider before installing

This skill appears to implement the described feedship→LLM→replace pipeline, but it contains a few practical and privacy-related issues you should consider before installing: - Hard-coded paths: SKILL.md uses /Users/y3/feedship and other local paths; update these to point to your feedship checkout or make them configurable. If you run it as-is, it will attempt to read that specific directory and may fail or read unexpected files. - OS assumptions: The date commands use BSD/macOS flags (date -v-1d). On Linux these will fail. Validate and adapt the commands for your environment. - Outbound publishing: The cron example posts the generated digest to channels/targets you provide. Double-check channel/destination parameters so reports are not published where you don't intend. - Local data access: The skill expects access to a local feedship project and will read article data from it. Ensure you trust the local repo and that no sensitive articles will be included unintentionally. - Testing: Run the pipeline manually in an isolated session first (use a small sample /tmp dataset) to confirm behavior, verify replace_refs.py output, and ensure the LLM prompt rules are respected. If the developer can confirm that /Users/y3 is just an example and provide instructions for parameterizing paths and cross-platform date handling, my confidence in the coherence would increase.

AI 日报 (Feedship AI Daily)

Version: 1.21.0 For: OpenClaw compatible agents Description: Generate daily AI news digest via feedship article extraction + AI strategic analysis

为什么需要引用替换（Step 3.5）

LLM 生成的引用有两大幻觉风险：

1. 编号幻觉：引用列表中不存在的编号（如 ${999}）
2. 展开幻觉：在引用中直接输出标题/链接时捏造内容

解决方案： LLM 只输出 ${N} 占位符，标题和链接由 replace_refs.py 脚本从权威 JSON 中注入。脚本对无效编号会输出 [无效引用 #N] 并打印警告。

---

1. Setup

This skill requires feedship v1.8.0+. Use the local project version:

cd /Users/y3/feedship && uv run feedship --version

Verify article command is available:

cd /Users/y3/feedship && uv run feedship article --help

2. Usage

On-Demand (Manual)

User triggers: "生成今日日报", "今日新闻摘要", "daily digest", "AI日报", "生成简报", "大模型日报"

The agent will activate this skill and run the Generate Daily Report flow (see section 3).

Automatic (Cron)

Schedule daily reports at 8:00 AM Beijing time.

openclaw cron add \
  --name "feedship-ai-daily" \
  --agent feedship-ai-daily \
  --cron "0 8 * * *" \
  --tz Asia/Shanghai \
  --session isolated \
  --announce \
  --channel <your-channel> \
  --to <your-destination> \
  --timeout-seconds 900 \
  --message "使用 feedship-ai-daily skill 生成今日日报。" \
  --thinking xhigh

---

3. Generate Daily Report

Step 1. 提取今日文章（带过滤）

# 提取今日文章
cd /Users/y3/feedship && uv run feedship article list \
  --since $(date -v-1d +%Y-%m-%d) \
  --until $(date +%Y-%m-%d) \
  --limit 3333 \
  --json > /tmp/today_articles_raw.json

# 过滤出 AI/技术相关标题（节省 token，提升信号比）
python3 -c "
import json, sys
data = json.load(open('/tmp/today_articles_raw.json'))
items = data.get('items', [])
keywords = 'AI|model|agent|LLM|GitHub|open.?source|developer|software|tech|coding|compute|GPU|inference|ML|neural|benchmark|API|protocol|framework|tool'
filtered = [item for item in items if any(k in (item.get('title','') + item.get('description','')) for k in keywords.split('|'))]
# 输出过滤后标题
for i, item in enumerate(filtered):
    t = item.get('title','').strip()
    if t:
        print(f'{i+1}. {t}')
# 保存过滤后数据供替换脚本使用
with open('/tmp/today_articles_filtered.json', 'w') as f:
    article_map = {}
    for i, item in enumerate(filtered):
        t = item.get('title','').strip()
        if t:
            article_map[str(i+1)] = {'title': t, 'link': item.get('link', '')}
    json.dump(article_map, f, ensure_ascii=False, indent=2)
print(f'过滤后: {len(filtered)} 篇 / 原始: {len(items)} 篇', file=sys.stderr)
" > /tmp/article_titles.txt

验证结果:

head -10 /tmp/article_titles.txt
echo "---"
python3 -c "import json; print(f'替换脚本使用 {len(json.load(open(\"/tmp/today_articles_filtered.json\")))} 篇文章')"

Step 2. 拼接提示词

从 references/prompt.md 读取提示词模板，拼接过滤后的标题列表：

ARTICLE_TITLES=$(cat /tmp/article_titles.txt)
PROMPT=$(cat skills/feedship-ai-daily/references/prompt.md)
# 拼接：将 # Input Data 后追加标题列表
FULL_PROMPT=$(echo "$PROMPT"; echo ""; echo "$ARTICLE_TITLES")

Step 3. 发送至 LLM

将 $FULL_PROMPT 发送给 LLM（MiniMax-M2.7 或同类模型）生成分析。

⚠️ 重要提示给 LLM： 在回复中，严格只使用 ${N} 格式引用新闻，禁止展开标题或链接。

Step 3.5. 替换引用占位符

TODAY_ARTICLES=$(python3 -c "import json; print(json.dumps(json.load(open('/tmp/today_articles_filtered.json'))))")
cat LLM_OUTPUT.txt | TODAY_ARTICLES="$TODAY_ARTICLES" \
  python3 skills/feedship-ai-daily/scripts/replace_refs.py > /tmp/daily_report_final.md

脚本功能：

• ${N} → [标题](链接)
• ${3,7} → 展开为两个独立链接（空格分隔）
• 无效编号 → [无效引用 #N]（带警告）
• 列表项格式 ${N}｜中文标题 → ${N} 被替换，保留 ｜中文标题
• 脚本底部打印统计：[replace_refs] N 处引用已替换

Step 4. 输出最终报告

读取 /tmp/daily_report_final.md 输出给用户。

---

4. Prompt 模板参考

完整的提示词模板位于 references/prompt.md，包含：

• Role: Principal Tech Strategist & Open-Source Trend Forecaster
• Citation Rules: 防幻觉核心规则——LLM 只输出 ${N}，禁止展开标题/链接
• Execution Steps: 4个步骤（分类解构 → 根本原因 → 暗线发现 → 价值翻译）
• Output Rules: 语气、结构、约束条件

---

5. Troubleshooting

Problem	Solution
feedship: command not found	使用 cd /Users/y3/feedship && uv run feedship
report command not found	全局版本过旧，使用本地 v1.8.0+ 版本
Empty article list	检查日期范围，确认有文章发布
大量 [无效引用 #N]	LLM 引用了不存在的编号，检查 prompt.md 中的 Citation Rules 是否被遵循
LLM 直接展开标题/链接	在 Step 3 发送时强调：禁止展开链接，只输出 ${N}
LLM timeout	减少 --limit 数量；或确保文章过滤后数量在 100 以内

---

6. Change Log

• v1.21.0: 新增文章过滤步骤（AI/Tech 关键词过滤）；强化 Citation Rules 防幻觉约束；替换脚本支持逗号分隔多引用和无效编号警告
• v1.20.0: 初始版本，从 feedship report 迁移到手动提取 → LLM → 替换 pipeline