search

name: search description: "Unified search & research. Automatically calibrates depth — from quick fact-checks to multi-source investigations — in a single adaptive flow. Triggers: /s, "search for", "look up", "verify", "any updates on", etc."

Search — Adaptive Research Skill

One flow, automatic convergence. No separate "quick" vs "deep" modes — depth is determined by the question itself. Stop when you have enough.

Tool Constraints

• Web search: Exa (bash skills/exa-search/scripts/search.sh "query" [count] [type] [category])
• Full-text extraction: web_fetch / bash skills/exa-search/scripts/content.sh "url1" "url2"
• Community search: bird search "query" -n count / bird read <id-or-url>
• 中文社媒：python3 skills/search/scripts/cn-social.py <platform> [options]（⚠️ 门控工具，见下方规则）
• Forbidden: browser

🚫 中文社媒工具门控规则（cn-social）

cn-social 覆盖知乎、B站、小红书、微信公众号、TikTok。仅在以下条件满足时才可调用：

1. 用户明确指定要搜索国内社交媒体（如"搜一下国内社媒"、"看看国内怎么说"）
2. 用户直接点名平台：知乎 / B站 / 小红书 / 微信公众号 / TikTok

不满足上述条件时，严禁使用 cn-social——即使是"全网调研"、"全网搜索"也不包含 cn-social。

触发后，cn-social 作为额外信源与 Exa/bird/web_fetch 并行使用，搜索流程规范不变。

平台与参数速查

可选参数

小红书搜索：

• --sort TYPE — 排序：general / popularity_descending / time_descending / comment_descending / collect_descending
• --note-type TYPE — 类型过滤：不限 / 视频笔记 / 普通笔记
• --note-time TIME — 时间过滤：不限 / 一天内 / 一周内 / 半年内
• --page N — 页码（默认1）
• --search-id ID / --session-id ID — 翻页标识（首次请求返回）

TikTok搜索：

• --offset N — 偏移量（默认0）
• --count N — 数量（默认10）

通用：--raw — 输出原始JSON

示例

# 知乎热榜
python3 scripts/cn-social.py zhihu-hot

# 知乎AI搜索（两步异步，自动轮询）
python3 scripts/cn-social.py zhihu-search --keyword "DeepSeek"

# 小红书搜索（带时间过滤）
python3 scripts/cn-social.py xiaohongshu --keyword "AI绘画" --sort time_descending --note-time 一周内

# 小红书笔记全文（先搜索获取 id，再拉详情）
python3 scripts/cn-social.py xiaohongshu-note --note-id 6990b4cf000000001a02770a

# B站视频详情
python3 scripts/cn-social.py bilibili --bvid BV1ttk9YkEVx

# 微信公众号文章
python3 scripts/cn-social.py wechat --ghid gh_4cd0897a7077

# TikTok搜索
python3 scripts/cn-social.py tiktok --keyword "AI" --count 5

已知限制

• 抖音（Douyin）：API 返回参数错误，不可用
• 微博（Weibo）：API 要求登录，实际无法获取数据
• B站搜索：302.AI 仅提供视频详情接口，无搜索接口
• 知乎AI搜索：异步两步流程，脚本自动轮询（约 4-16 秒），返回 AI 摘要 + 相关内容卡片
• 小红书搜索：必须用 V3 app 路径（脚本已内置）
• 302.AI 官方声明：信息搜索系列接口无法保障稳定性，仅限个人体验使用

API 文档

• 单接口 OpenAPI 文档：https://doc.302.ai/{api_id}e0.md
• 全量 API 列表：https://doc.302.ai/llms.txt

Output Rules

• Deliver exactly one message to the user: the final report
• All intermediate steps are silent — no progress spam
• Complex research can be archived to memory/research/YYYY-MM-DD_<slug>/

Pipeline

Stage 0: Intent Analysis (silent)

On receiving a task, complete the following internally — do not output to the user:

1) Task Classification:

2) Search Element Decomposition:

• Core entities, time window, type of change
• Web search keywords (official/general)
• X search keywords (community/experience/screenshots)

Proceed directly to Stage 1.

Stage 1: Recon

Use 1–2 lightweight searches to establish a knowledge frame:

1. Exa broad search (bash skills/exa-search/scripts/search.sh "<core query>" 10 auto)
2. Internal notes: canonical names, core keywords, source directions
3. Evaluate: Does this search already answer the user's question?
   • Simple fact with a reliable source → skip to Report
   • Information gaps remain → proceed to Stage 2

Stage 2: Evidence Collection Loop

Iterative loop, each round:

1. Identify the biggest information gap (What's missing? Which claims lack cross-verification?)
2. Search targeting the gap:
   • Need official/general sources → Exa search
   • Need community opinions/rumors → bird search
   • Need to verify a specific URL → web_fetch
   • Only when user explicitly requests Chinese social media → cn-social.py
3. web_fetch key sources for full-text verification
4. Update internal evidence map, reassess coverage

Stop conditions (stop when any is met):

• ✅ Core question has a clear answer supported by reliable sources
• ✅ Multiple independent sources cross-confirm key conclusions
• ✅ Source strategy fulfilled (per Stage 0 classification)
• ⚠️ Two consecutive rounds yielded no new useful information (diminishing returns)
• ⚠️ Genuinely nothing found (report honestly, describe attempted search paths)

No hard round limit. Simple questions converge in 1–2 rounds; complex ones may take 4–6. The model decides.

Stage 3: Source Grading

Tag all collected sources with credibility:

• 🟢 High: Official sources, authoritative media first-hand reports, official docs
• 🟡 Medium: Reputable media second-hand reports, industry KOLs, verifiable community info
• 🔴 Low: Anonymous sources, unconfirmed rumors, single community posts

Present conflicting information with different claims and their respective sources. Do not take sides.

Report Format

📌 Conclusion
One-line core answer + necessary caveats [¹](URL)

📋 Details
- Organized by logic, not search order
- Key conclusions have inline footnote citations [¹](URL) for direct verification
- Common-sense transitions and reasoning need no citation
- Conflicting claims listed with sources for each side

📎 Core Sources
1. [Title/Description](URL) 🟢/🟡/🔴
2. ...

Citation Rules

• Inline: Attach [¹](URL) footnotes after key facts, data, and core conclusions. Numbers correspond to the source list at the end
• Source list: A superset of inline citations — includes both directly cited sources ([¹] [²]...) and high-quality sources used for cross-verification but not directly referenced in text
• Sort by importance/credibility, tag 🟢🟡🔴

Format flexibility: Simple questions can be short (conclusion + 1–2 sources); complex questions will naturally be longer. Length is driven by content, not template padding.

Source Usage Rules

See references/source-priority.md and references/community-gates.md for details.

Core principles:

• Anchor conclusions on official/authoritative sources; use community info for leads and anomaly detection
• Conclusions based solely on community sources must be tagged "unconfirmed by official sources" with reduced confidence
• Key information requires at least two independent sources for cross-verification; single-source claims must be explicitly noted

Edge Cases

• Nothing found: Report honestly, describe the search paths attempted. Never fabricate content
• Staleness risk: Note when information was retrieved
• Controversial/sensitive topics: Present multiple perspectives without value judgments
• Time-sensitive queries: Prioritize time-bounded search terms (year/month/"launch"/"release notes"); broaden only if nothing is found

Triggers

• /s "query" — unified search entry point
• Natural language: "search for", "look up", "verify", "any updates on", etc.