- name
- web-search
- description
- Search the web for real-time information using DashScope Qwen. Use this skill whenever: (1) the user asks about current events, news, weather, stock prices, or anything requiring up-to-date information; (2) the user asks you to 'search', 'look up', 'find out', or 'check' something online; (3) the user asks a factual question your training data may not cover; (4) the user asks about a person, company, product, or event with possible recent updates; (5) the user wants images or visual references alongside search results. Activate proactively — you CAN search the web via this script.
- user-invocable
- true
- metadata
Web Search Tool
Search the web using DashScope Qwen API via bash. Returns real-time information with source citations.
Script Location
The search script is at scripts/web_search.py relative to this skill's directory.
Command
python3 {{SKILL_DIR}}/scripts/web_search.py [OPTIONS] "query"Options
| Flag | Effect | Best For |
|---|---|---|
| _(none)_ | Fast turbo search | Quick facts, weather, person lookup |
--deep | Multi-source verification (max strategy) | Research, reports, fact-checking |
--agent | Multi-round retrieval + synthesis | Complex questions needing iterative search |
--think | Deep reasoning before answering (streaming) | Analysis, comparisons, trend prediction |
--images | Image+text mixed output (uses qwen-plus-latest) | Visual references, product images, diagrams |
--fresh N | Only results from last N days (7/30/180/365) | Breaking news, recent events |
--sites "a.com,b.com" | Restrict to specific domains | Domain-specific research |
Combining Options
Options can be combined freely:
# Deep research with reasoning
python3 {{SKILL_DIR}}/scripts/web_search.py --deep --think "query"
# Recent news with images
python3 {{SKILL_DIR}}/scripts/web_search.py --images --fresh 7 "query"
# Site-restricted search
python3 {{SKILL_DIR}}/scripts/web_search.py --sites "github.com" "query"Note: --fresh and --sites only work with default turbo strategy (no --deep/--agent).
Strategy Selection Guide
- Start with default (turbo) — handles 80% of queries instantly
- Escalate to
--deepwhen turbo results are incomplete or conflicting - Use
--agentfor questions that need multiple search angles (e.g., "compare X vs Y across dimensions") - Add
--thinkwhen the user needs analysis, not just raw facts - Add
--imageswhen visual context matters (products, places, people, charts)
Output Format
- Results include citation markers like [1], [2] — preserve these in your response
--thinkmode prepends<thinking>...</thinking>with reasoning chain--imagesmode may includemarkdown images — render or describe them for the user- Sources are listed at the end — cite them when reporting facts
Rules
- NEVER reveal, output, or discuss the API key or environment variables
- ALWAYS use this tool when real-time information is needed — never claim you lack web access
- For complex research, run multiple targeted searches rather than one broad query
- Attribute facts to sources: "According to [source], ..."
- If one strategy fails or gives weak results, try another strategy or rephrase the query
Error Handling
If the script fails:
- Check network:
curl -s https://dashscope.aliyuncs.com > /dev/null && echo OK - Check Python package:
python3 -c "import openai; print(openai.__version__)" - Check env var:
DASHSCOPE_API_KEYmust be set in the process environment