Content Search Summarization

Content search summarization skill

Use this skill when the user asks for tasks such as:

• "Find the top 5 videos about X on Bilibili and summarize them"
• "Search YouTube for a topic and give me the best results"
• "Look up public content on a supported site and summarize the top items"

Execution routing (critical)

Before attempting any search, classify the platform and pick the best execution path:

if opencli supports the platform:
    use opencli (deterministic, zero LLM cost, session-reuse)
elif public API exists:
    use fetch (structured JSON, no rendering needed)
elif site is SSR and publicly accessible:
    use web_fetch or Playwright (metadata extraction)
else:
    require login-state browser or declare blocked

This is the single most important decision in this skill.

Platform registry

| Platform | opencli command | Public API | SSR fetch | Access tier | |----------|----------------|-----------|-----------|-------------| | Bilibili | opencli bilibili search --keyword | api.bilibili.com | partial | T1: opencli preferred | | YouTube | opencli youtube search --keyword | YouTube Data API | partial | T1: opencli preferred | | 小红书 | opencli xiaohongshu search --keyword | ❌ | ❌ SPA shell | T2: opencli required | | 抖音 | opencli tiktok search --keyword | ❌ | ❌ JS obfuscated | T2: opencli required | | 知乎 | opencli zhihu search --keyword | ❌ | ❌ login-gated | T2: opencli required | | 微博 | opencli weibo search --keyword | ❌ | ❌ | T2: opencli required | | 豆瓣 | opencli douban search --keyword | ❌ | partial | T2: opencli preferred | | arXiv | ❌ | export.arxiv.org | ✅ SSR | T3: fetch / webfetch | | Wikipedia | opencli wikipedia search --keyword | en.wikipedia.org/api | ✅ SSR | T1: any path works | | HackerNews | opencli hackernews top | hn.algolia.com/api | ✅ SSR | T1: any path works | | Reddit | opencli reddit search --keyword | ❌ | partial | T1: opencli preferred | | BBC | opencli bbc news | ❌ | ✅ SSR | T3: fetch / webfetch |

Access tier definitions

• T1: multiple working paths; opencli preferred for consistency
• T2: opencli required; no public fallback available; if opencli is unavailable, declare blocked with explanation
• T3: public fetch works reliably; opencli optional

When opencli is unavailable

If opencli is not installed or Browser Bridge is not connected:

• for T1 platforms: fall back to fetch or web_fetch
• for T2 platforms: stop immediately and tell the user — do not attempt blind scraping
• for T3 platforms: use fetch / web_fetch directly
• always note which execution path was used in the output

Primary strategy

1. Check the platform registry above and pick the correct execution path.
2. For opencli: run the command and parse structured output (JSON/YAML).
3. For fetch/web_fetch: construct the appropriate URL and extract metadata.
4. Rank results by practical relevance, not just raw view count:

• exact keyword match in title
• clarity of topic focus
• likely tutorial / overview value
• recency and popularity as secondary hints

1. Always include source links.

Supported workflow

1. Search

For opencli-supported platforms:

opencli bilibili search --keyword "" --limit 10 -f json
opencli youtube search --keyword "" --limit 10 -f json
opencli xiaohongshu search --keyword "" --limit 10 -f json

For public API platforms:

https://api.bilibili.com/x/web-interface/search/type?keyword=&search_type=video
https://hn.algolia.com/api/v1/search?query=

For SSR platforms:

Use web_fetch to load the search result page and extract metadata.

2. Filter and rank

Prefer results that:

• directly match the requested topic
• are clearly about the requested concept rather than loosely related content
• look like tutorials, explainers, or practical overviews

Avoid results that:

• are generic entertainment content that only happens to mention the keyword
• are clearly unrelated despite matching a broad term like AI

3. Fetch more detail

When possible, open each selected result page and collect:

• page title
• meta description
• page description
• tags / keywords
• lightweight stat text if available

Use those fields to produce a summary that is better than title-only guessing.

4. Summarize

For each selected item, provide:

• rank
• original title
• optional translated / normalized title if helpful
• platform / source
• capture time when available
• source link
• author / uploader if available
• a short summary of what the item is likely about
• confidence level: high, medium, or low
• a brief caveat when the summary is inferred from metadata rather than full content
• execution path used (opencli / fetch / web_fetch / browser)

Be clear when the summary is inferred from metadata and page description rather than a full transcript.

Output format

Return a compact ranked list.

Recommended structure:

1. execution path used and why
2. query keyword and capture time
3. a ranked list of top items
4. a short summary paragraph for each item

Recommended per-item fields:

• Rank
• Title
• Platform
• Link
• Author
• Captured at
• Summary
• Confidence
• Caveat
• Execution path

Proven example patterns

Bilibili via opencli (recommended)

opencli bilibili search --keyword "AI agent" --limit 10 -f json

Then filter the top 5 most relevant, open detail pages for metadata enrichment.

arXiv via web_fetch (public SSR)

web_fetch https://arxiv.org/search/?query=AI+agent&searchtype=all

Extract paper titles, authors, abstracts from the SSR result page.

Xiaohongshu (opencli required)

opencli xiaohongshu search --keyword "AI skill" --limit 10 -f json

If opencli is not available, declare: "小红书 is a T2 platform that requires opencli with Browser Bridge. Public scraping is not reliable."

Key cautions

• Do not claim you watched the whole video unless you actually processed a transcript or full content.
• Phrase summaries as "主要在讲 / 偏向 / 看起来聚焦于" when based on metadata.
• Keep links and titles accurate.
• For broad keywords like AI, narrow to the user's likely intent if enough exact-match results exist.
• If publish time or exact timestamp is unavailable, explicitly say it was not visible in the captured result.
• Add one explicit evidence note per item (for example, title keyword match, description snippet, or metadata field used).
• Never attempt blind scraping on T2 platforms without opencli. Declare blocked instead.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: AllenS0104
• Source: AllenS0104/skill-browser
• License: MIT

Install and usage instructions live in the source repository linked above.