← Back to Skills Marketplace
yanpeipan

Ai Daily

by Yan Zer0 · GitHub ↗ · v1.21.0 · MIT-0
cross-platform ⚠ suspicious
192
Downloads
0
Stars
0
Active Installs
12
Versions
Install in OpenClaw
/install feedship-ai-daily
Description
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...
README (SKILL.md)

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 \x3Cyour-channel> \
  --to \x3Cyour-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

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

$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
Usage Guidance
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.
Capability Analysis
Type: OpenClaw Skill Name: feedship-ai-daily Version: 1.21.0 The skill exhibits high-risk behaviors including direct shell execution and file system manipulation to process untrusted external data (RSS feeds). It contains hardcoded absolute paths (e.g., `/Users/y3/feedship` in `SKILL.md`) and constructs shell commands using variables derived from article titles, which could lead to command injection or execution errors if the agent environment is not strictly isolated. While the workflow for generating AI news digests and the `replace_refs.py` script appear functionally aligned with the stated purpose, these architectural vulnerabilities and the use of broad permissions warrant a suspicious classification.
Capability Assessment
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.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install feedship-ai-daily
  3. After installation, invoke the skill by name or use /feedship-ai-daily
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.21.0
feedship-ai-daily v1.13.1 - Added script for citation placeholder replacement: `scripts/replace_refs.py` - Added prompt template reference: `references/prompt.md` - Removed old format instructions: `REPORT_FORMAT.md` - Workflow now emphasizes citation placeholder strategy and post-processing for references - Strengthened anti-hallucination rules for LLM-generated reports and clarified output steps
v1.13.0
feedship-ai-daily v1.13.0 (major update) - Expanded daily digest from 4 to 6 sections: now includes 政策解读 (Policy analysis) and 媒体热点 (Media trends). - Updated SKILL.md with a simpler, more step-by-step setup and usage guide. - Clarified instructions for automated (cron) and on-demand usage, with explicit command examples. - Enforced systematic processing and accurate article linking in all sections; AI generation of new content is prohibited. - Removed outdated files and all section content must use true article links from JSON search results.
v1.10.0
- Version updated from 1.8.0 to 1.9.0 in SKILL.md. - No changes observed to code or major instructions; only the version string in SKILL.md was updated. - All usage, setup, and report generation instructions remain unchanged.
v1.9.0
- Increased default cron timeout for scheduled reports from 600 seconds to 900 seconds. - Updated all example `openclaw cron add` commands to use the new 900-second timeout.
v1.8.0
- Updated documentation: modified IDENTITY.md and REPORT_FORMAT.md for clarity or improvements. - Removed obsolete file: BOOTSTRAP.md. - No changes to core functionality or usage instructions. - Version remains 1.7.0 in SKILL.md.
v1.7.1
No user-facing changes. SKILL.md version updated from 1.3 to 1.7.0 for metadata consistency.
v1.7.0
feedship-ai-daily v1.7.0 - Added system documentation files: AGENTS.md, BOOTSTRAP.md, HEARTBEAT.md, IDENTITY.md, SOUL.md, TOOLS.md, USER.md. - Updated setup and usage instructions in SKILL.md, emphasizing the required --message flag in cron jobs for proper skill file loading. - Enhanced OpenClaw cron job examples to include skill instruction messages for both Telegram and Feishu channels. - Clarified the report generation procedure and troubleshooting steps.
v1.6.0
**feedship-ai-daily v1.6.0 Changelog** - Major report format update: now outputs a 4-section digest (AI五层蛋糕, 精选推荐, 创业信号, 创作点), instead of the previous 6 sections. - Featured picks and hot topics are now merged into a single section. - Added REPORT_FORMAT.md documenting the new report specification. - Updated documentation and instructions in SKILL.md to reflect the new workflow and simplified structure.
v1.5.0
**feedship-ai-daily 1.5.0 introduces a new mandatory 6-section report format and enhances integration setup guidance.** - Digest reports now must include exactly 6 distinct sections (A–F): 按来源分组, AI摘要, 精选推荐, 热点话题, 创业信号, 创作点. - Strict output order and section presence is enforced—no skipping or combining allowed. - Expanded setup instructions for Telegram and Feishu delivery channels with detailed step-by-step guides. - Clarified OpenClaw automation, channel configuration, and testing steps for scheduled and on-demand use. - Previous 3-section format replaced; new documentation reflects the updated reporting flow.
v1.0.2
- No code or documentation changes detected in this version. - Functionality and instructions remain the same as previous version.
v1.0.1
- Updated system requirements to require feedship with cloudflare and ml extras. - Installation instructions now recommend: pipx install 'feedship[cloudflare,ml]' or uv pip install 'feedship[cloudflare,ml]'. - No other content or usage changes.
v1.0.0
- Initial release of feedship-ai-daily skill. - Automatically generates a daily AI news digest from your feedship subscriptions. - Digest includes three sections: Today’s new articles with summaries, hot topics clustering, and featured picks by feed weight. - Supports both on-demand generation and automated daily cron scheduling. - Requires feedship and uv; install with `uv pip install feedship`.
Metadata
Slug feedship-ai-daily
Version 1.21.0
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 12
Frequently Asked Questions

What is 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... It is an AI Agent Skill for Claude Code / OpenClaw, with 192 downloads so far.

How do I install Ai Daily?

Run "/install feedship-ai-daily" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Ai Daily free?

Yes, Ai Daily is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does Ai Daily support?

Ai Daily is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Ai Daily?

It is built and maintained by Yan Zer0 (@yanpeipan); the current version is v1.21.0.

💬 Comments