Search & Fetch

General search and content extraction skill, providing two core capabilities: Search Mode (multi-source parallel search + result synthesis) and Fetch Mode (URL content extraction + structured analysis).

Tool Overview

This skill orchestrates the following types of tools, combining them as needed:

Category	Tool	Description
Built-in Search	`WebSearch`	DuckDuckGo search, always available
MCP Search	Available MCP search tools in the current environment	Chinese search, professional search, etc., automatically selected based on availability
Technical Docs	`ctx7` CLI	Query official docs for programming libraries/frameworks
Browser	`agent-browser` CLI	Access search engines, extract content from JS-heavy pages (Rust native, token-efficient)
Content Extraction	`defuddle` CLI → `WebFetch` → MCP fetch → `agent-browser`	Degrade by priority

Refer to

references/tool-catalog.md

for specific invocation syntax.

Intent Detection

This skill is responsible for identifying and routing all URL and search intents. Automatically judge based on user input:

URL Detection

Scenario	Handling
Message contains only URL (no other text, or only URL + spaces/punctuation)	Directly enter Fetch Mode
Message contains URL + reading intent ("read this", "analyze", "take a look at this")	Directly enter Fetch Mode
Message contains URL + save intent ("save", "clip", "save", "archive")	After completing Fetch Mode, automatically trigger the Clipping Save of the knowledge base skill
URL appears in a long message with unclear intent	Ask: "Would you like me to extract and analyze the content, or just keep the link?"
URL is obviously tool-based (localhost, internal network address, file path)	Do not trigger

Search Detection

Scenario	Handling
User asks questions, researches topics, needs to find information	Search Mode
Provides URL + "search for more similar content"	First Fetch then Search
Delegated call by other skills	Execute according to the mode specified by the delegator

Search Mode

Multi-source parallel search, output structured results after deduplication and synthesis.

Step 1 — Query Classification

Judge the type based on query content, select tool combinations:

Query Type	Judgment Basis	Tool Combination
Technical Docs	Involves programming libraries, APIs, framework usage	`ctx7` (library → docs) + `WebSearch`
Chinese Content	Chinese query, or China-specific topics	MCP search + `WebSearch` in parallel
General Query	Other general questions	`WebSearch` + MCP search in parallel
In-depth Research	User explicitly requests comprehensive/in-depth research	All tools in parallel, including `agent-browser` to access Google/Bing

Step 2 — Parallel Search Execution

Based on classification results, call multiple search tools in parallel within the same response round. Refer to

references/tool-catalog.md

for tool invocation details.

Key Principles:

Call independent tools in parallel, no waiting
context7 has sequential dependencies: first
```
ctx7 library
```
to resolve library ID, then
```
ctx7 docs
```
to query content (max 3 calls per question)
Adjust wording for different search engines for the same query (e.g., send original Chinese query to MCP search, translate to English for WebSearch)
For in-depth research, use
```
agent-browser
```
to open Google/Bing search pages, extract results after getting element references via
```
snapshot
```

Step 3 — Result Synthesis

Deduplication: Deduplicate by URL, merge information from the same page from different sources
Grouping: Group by sub-topic or relevance
Sorting: Sort by relevance and information quality
Output Format:

markdown

## Search Results: [Query Topic]

### [Sub-topic 1]
- Key point summary (Source: [Title](url))
- Key point summary (Source: [Title](url))

### [Sub-topic 2]
- ...

---
**Source List**
1. [Title](url) — Brief description
2. [Title](url) — Brief description

Step 4 — Follow-up Operations

After displaying search results, provide follow-up options:

"Would you like me to read the full content of a certain link?" → Switch to Fetch Mode
"Would you like to save to the knowledge base?" → Guide users to trigger knowledge management skills like obsidian-knowledge
Users can continue to ask questions to narrow or expand the search scope

Fetch Mode

Extract clean content from URLs and perform structured analysis.

Step 1 — Content Extraction

Attempt in priority order, automatically degrade if the previous one fails (error, returns empty, content < 100 characters):

1. defuddle parse <url> --md          ← Preferred, removes noise
2. WebFetch(url)                       ← General degradation
3. MCP fetch tool (if available)               ← MCP degradation
4. agent-browser open <url> → snapshot → get text   ← Final resort for JS-heavy pages

Record which tool was used after successful extraction, to transparently inform users.

Refer to

references/tool-catalog.md

for tool invocation details.

Step 2 — Structured Analysis

Selectively apply analysis dimensions based on content type. See

references/analysis-template.md

for detailed templates.

Content Type	Applicable Dimensions
Articles, comments, papers	All five dimensions (text deconstruction → concept extraction → critical examination → idea deepening → practical transformation)
News, product pages	Text deconstruction + concept extraction + practical transformation
Tutorials, technical docs	Text deconstruction + concept extraction + practical transformation
Data pages, API docs, reference tables	Skip analysis, directly present extracted content

Step 3 — Output

Two output modes:

Standalone Mode (when used directly by users):

markdown

## [Page Title](url)

**Author**: xxx | **Date**: xxx | **Word Count**: xxx

### Summary
A paragraph summary.

### Analysis
[Output applicable dimensions according to analysis-template.md]

### Key Points
- Point 1
- Point 2
- Point 3

After analysis output, actively prompt: "Would you like to save to the knowledge base?" If the user agrees, guide to trigger the Clipping Save process of obsidian-knowledge.

Integration Mode (when delegated by other skills):

Output the following information blocks in natural language, for direct consumption by the caller (do not output JSON):

Page Title and Original URL
One-sentence Summary
Key Points (list of points)
Five-dimensional Analysis (output applicable dimensions)
Full Clean Text (extracted clean markdown full text)
Extracted Entities: Mentioned tools/software, stocks/companies, people (name + identity), books/papers/resources

Degradation Strategy

Handling methods when a tool is unavailable:

Tool Unavailable	Impact	Alternative Solution
defuddle	Cannot extract clean main content	Degrade to WebFetch
WebFetch	Reduced extraction capability	Use MCP fetch or agent-browser
MCP search	Reduced search sources	Use WebSearch for all queries
ctx7	No library doc search	Use WebSearch + site:xxx targeted search
agent-browser	Cannot process JS-heavy pages	Inform users, use WebFetch to extract as much as possible
All search tools	Cannot search	Inform users, suggest providing URL directly

Do not interrupt the process because a tool fails — always try the next degradation solution until all are exhausted before informing the user.

Integration Protocol

This skill can be called by other skills. The delegator only needs to specify "delegate to search-and-fetch" in their own SKILL.md.

Behavior When Delegated

Execute according to the mode (Search/Fetch) specified by the delegator
Use Integration Mode output format
Do not directly interact with users about follow-up questions (handled by the delegator)
If content extraction completely fails, return failure marker and reason, and the delegator decides how to inform the user

search-and-fetch

NPX Install

Tags

SKILL.md Content (Chinese)