Multi-Source Intel Radar
Create a founder-grade signal system: less junk, more decisions.
Inputs
- - OPML file (default:
/Users/rogeryang/Downloads/follow.opml) - Keyword whitelist (default: 创业, AI, 增长, 金融)
- Optional source lists for non-RSS channels (X list links, subreddit list, WeChat/XHS accounts)
Output Contract
Always output:
- 1. Top 3 must-read signals (one-line why + one action + clickable source link)
- Top 5 watchlist items (with source link)
- Dropped noise summary (what got filtered and why)
- Filter transparency (counts + rates: scanned -> matched -> shortlisted -> top3)
- Next experiment (one concrete growth/ops move)
Workflow
Step 1) Ingest feed sources
- - Parse OPML with INLINECODE1
- Generate normalized feed list: INLINECODE2
Step 2) Fetch + filter
- - Run
scripts/build_digest.py with keyword whitelist - Time window default: last 48h
- Keep only items that match whitelist in title/summary
- For Xiaohongshu: do browser search (not watchlist-dependent), using keyword combos like:
- 创业 AI 增长
- AI 产品 复盘
- 增长运营 案例
Then append top findings with profile/note links.
Step 3) Score items
Use this weighted scoring:
- - Relevance to whitelist (40%)
- Actionability in 7 days (30%)
- Novelty / non-obviousness (20%)
- Evidence density (10%)
Step 4) Summarize for execution
For each selected item, provide:
- - Core insight (1 sentence)
- Why it matters for current product
- Suggested action today (1 step)
Source Coverage Notes
- - YouTube/X/Reddit often available via RSSHub or platform feeds in OPML
- WeChat OA and Xiaohongshu are often not natively RSS; add via:
- RSS bridge links (if available)
- Manual watchlist files (
assets/wechat_watchlist.txt,
assets/xhs_watchlist.txt)
- - If a source has no feed, include it in watchlist and mark as manual scan required
Anti-Noise Rules
- - Do not output generic motivational posts
- Drop repeated观点 without new evidence
- Prefer first-hand data / concrete case over opinion
- Keep digest under 10 items total
Daily Cadence (recommended)
- - 09:30: morning digest (strategic)
- 18:30: evening digest (tactical)
Keyword Defaults
创业, AI, 增长, 金融
If user provides new keywords, merge and deduplicate.
Files
- - Parser: INLINECODE6
- Digest builder: INLINECODE7
- Notes: INLINECODE8
多源情报雷达
打造创始人级信号系统:更少噪音,更多决策。
输入
- - OPML文件(默认:/Users/rogeryang/Downloads/follow.opml)
- 关键词白名单(默认:创业、AI、增长、金融)
- 非RSS频道的可选来源列表(X平台列表链接、Reddit子版块列表、微信/小红书账号)
输出约定
始终输出:
- 1. 前3条必读信号(一行原因 + 一个行动 + 可点击来源链接)
- 前5条观察清单项目(含来源链接)
- 已过滤噪音摘要(被过滤内容及原因)
- 过滤透明度(数量与比率:扫描数 -> 匹配数 -> 入围数 -> 前三数)
- 下一步实验(一个具体的增长/运营动作)
工作流程
步骤1)导入订阅源
- - 使用 scripts/parse_opml.py 解析OPML
- 生成标准化订阅源列表:assets/feeds.txt
步骤2)获取并过滤
- - 使用关键词白名单运行 scripts/build_digest.py
- 时间窗口默认:最近48小时
- 仅保留标题/摘要中匹配白名单的项目
- 针对小红书:执行浏览器搜索(不依赖观察清单),使用关键词组合如:
- 创业 AI 增长
- AI 产品 复盘
- 增长运营 案例
然后附加含个人主页/笔记链接的顶级发现。
步骤3)评分项目
使用加权评分标准:
- - 与白名单相关性(40%)
- 7天内可操作性(30%)
- 新颖性/非显而易见性(20%)
- 证据密度(10%)
步骤4)总结为执行方案
为每个选定项目提供:
- - 核心洞察(一句话)
- 对当前产品的重要性
- 今日建议行动(一个步骤)
来源覆盖说明
- - YouTube/X/Reddit通常可通过OPML中的RSSHub或平台订阅源获取
- 微信公众号和小红书通常无原生RSS;通过以下方式添加:
- RSS桥接链接(如有)
- 手动观察清单文件(assets/wechat
watchlist.txt、assets/xhswatchlist.txt)
- - 若来源无订阅源,将其纳入观察清单并标记为需手动扫描
防噪音规则
- - 不输出通用励志帖
- 丢弃无新证据的重复观点
- 优先选择一手数据/具体案例而非观点
- 摘要总数控制在10条以内
每日节奏(推荐)
- - 09:30:早间摘要(战略层)
- 18:30:傍晚摘要(战术层)
关键词默认值
创业、AI、增长、金融
若用户提供新关键词,合并并去重。
文件
- - 解析器:scripts/parseopml.py
- 摘要构建器:scripts/builddigest.py
- 备注:references/scoring-and-ops.md