| name | web-lookup |
| description | Use when the user asks about something the robot doesn't know from its local knowledge — facts, current events, how-to questions, product specs, news, definitions, or anything requiring up-to-date information from the internet. Also triggers on "search for", "look up", "find information about", "research", "what is the latest on", "google", "what does X mean", "how do I", "specs for".
|
| version | 1.1 |
| requires | [] |
| consent | none |
| tools | ["web_search"] |
| max_iterations | 3 |
Web Lookup Skill
Use this skill to answer questions that require current or external information.
Steps
- Call
web_search(query) with a concise, targeted search query derived from the user's question
- Review the returned results (title + snippet)
- Synthesise a clear, grounded answer citing the most relevant result
- If results are poor quality, try one follow-up search with a refined query
Guidelines
- Keep queries short and specific — avoid full sentences
- Prefer authoritative sources (official docs, Wikipedia, reputable news)
- If no useful results found: say so honestly, don't fabricate
- Do NOT use this skill for questions about the robot's own sensors or status — use
get_telemetry for those
- Cite your source: "According to [title]..."
References
See references/query-patterns.md for query shaping tips and examples.
Gotchas
- Search ≠ browse —
web_search returns snippets only, not full page content; if the snippet is insufficient for a factual answer, say "I found a reference but can't confirm details without browsing — here's what I found: [snippet]"
- Stale results — search results can be 6–24h old; for truly real-time data (stock prices, live events) caveat with "as of [result date if available]"
- Robotics specs — for hardware like Feetech STS3215, LeRobot, OAK-D: the official docs and GitHub READMEs are the most reliable; avoid third-party summaries which are often outdated
- Don't over-search — one good query + follow-up is usually enough; 3 failed searches probably means the information isn't publicly available, not that the query needs more tries
- Robot-specific questions — "what is OpenCastor?" or "how does RCAN work?" should use local knowledge (trajectory/skill memory) first, not web search; web search is for external world knowledge
Example
User: "What is a Feetech STS3215 servo?"
→ web_search("Feetech STS3215 servo specifications")
→ Summarise: torque, voltage, protocol, use cases
User: "What's the latest news on robot manipulation research?"
→ web_search("robot manipulation research 2026")
→ Synthesise top 2–3 results into a summary
