| name | source-intel-recommender |
| description | Recommend high-depth sources for broad and edge-case learning (including adversarial SEO/security perspectives), using Karpathy's 92 feeds and curated deep GitHub source maps. Use when user asks for 信源推荐, 边缘玩法研究入口, 深水论坛/论文/情报入口, or wants predicted knowledge value. |
Source Intel Recommender
When To Use
Use this skill when the user asks:
- 针对某个问题推荐高价值信源
- 想要更广视野,尤其是边缘/对抗思路
- 先看什么、后看什么最有效
- 这些来源可能带来什么知识价值
Inputs
- User question and constraints (time, language, risk appetite, build stage)
references/karpathy-92-hn-2025.tsv
references/deep-sources-github.tsv
references/deep-sources-telegram.tsv (optional dynamic pool)
Workflow
- Detect query focus tags (SEO/PSEO/GEO, adversarial, OSINT, security, research, automation, commerce).
- Build candidate pool:
- Karpathy 92 long-form feed list (depth baseline)
- Deep GitHub source maps (entry points for high-signal communities and docs)
- Telegram channel pool (fast-moving adversarial/community signals)
- Score each source on 1-5:
Relevance: fit to query intent
Depth: insight density / first-principles value
Actionability: can turn into testable workflow quickly
Freshness: likely still useful now
Consequence: potential downside if copied blindly (higher score = better controllability)
- Compute
KnowledgeValue:
KnowledgeValue = 0.30*Relevance + 0.25*Depth + 0.25*Actionability + 0.10*Freshness + 0.10*Consequence
- Output top recommendations:
- Why source fits
- What likely value it unlocks
- Reading order (fast execution -> deep method -> adversarial counterexample)
- Potential consequences (account risk, ranking volatility, reputation, legal exposure) when relevant
- Missing list types needed to further expand deep-zone coverage
Output Format
问题标签:
推荐信源 Top N:
<source> | KnowledgeValue: x.x/5 | 预计知识价值:
- ...
优先阅读顺序: <3-5 items>
可能后果: <only when query has edge/adversarial intent>
建议补充列表类型: <1-5 precise categories>
实验建议: <1 action in 24h>
Heuristics
- Prefer high-signal entry maps over random tactic posts.
- For adversarial topics, explain likely consequences instead of only binary prohibition wording.
- Mix at least one deep methodology source and one executable source map.