Web Search
This skill uses
— a command-line search tool that returns DuckDuckGo results as plain text or JSON. It's built for LLM agents: plain text output is compact (title, URL, snippet per result) and the flags let you control time, region, site, and count.
When to reach for this
- The user asks about something recent ("latest", "today", "current", a year at or past your knowledge cutoff)
- A claim needs verification
- Documentation, references, or external context is needed
- The user asks to search, look up, find out, research, or check something
- You're answering a factual question and you're not sure your knowledge is current
Once you're already in a conversation, you don't need explicit permission — just search.
Command
The CLI is invoked as
(a short alias for
).
sh
ddg rust async tutorial # plain text, default 10 results
ddg -n 5 "best Rust web framework 2026"
ddg --json -n 5 "topic" # JSON envelope, for parsing
echo "tokio runtime overhead" | ddg # query via stdin
Plain text is the default and the right choice for most uses — it's compact and you can quote URLs and snippets straight out of it.
Flags worth knowing
— number of results (1–100, default 10). Use 3–5 for quick lookups, 10–20 for broader research. More results cost tokens and the long tail rarely helps.
— restrict to past day / week / month / year.
Critical for "latest" or "recent" questions — without it, evergreen pages outrank fresh ones and you'll quote stale info.
— repeatable site filter. Use for official docs (
), known-trusted sources, or to dodge spam.
— region (e.g.,
,
,
,
). Match the language of the query. Run
to see them all.
— compact JSON envelope. Use only when parsing programmatically (counting results, extracting URLs into a script). For ordinary reading and citing, plain text is better.
— paginate when the first page didn't have what you needed.
Patterns
Quick fact lookup
sh
ddg -n 3 "current ceo of openai"
Three results is enough to triangulate a consensus answer.
Latest news / recent updates
sh
ddg -n 5 -t w "anthropic claude release"
The
filter is what makes this useful — without it you'll get older articles that rank well.
Documentation
sh
ddg -n 5 -w docs.python.org "asyncio gather exception"
Fact-check a specific claim
sh
ddg -n 5 "the exact phrasing of the claim"
Read snippets carefully. If sources disagree, surface that to the user — don't pick one and present it as fact.
Broader research
Run several focused queries rather than one giant one. Three
queries on different angles beat one
.
Reading the output
Each result has a position, title, URL, and a short snippet. The snippet is usually enough to decide whether to fetch the page in full with
. When citing to the user, include the URL so they can verify; if you only have a snippet, say "from the search snippet..." rather than implying you read the whole page.
Gotchas
- Rate limits: in mode the response includes
meta.rate_limit.tokens_remaining
- No instant answer is normal — most queries just return the numbered list.
- Region must match language — a Korean query with will return weaker results; use .
- Don't pipe to to truncate — use instead, otherwise a result may get chopped mid-snippet.
- If the binary isn't installed, install with or
cargo install duckduckgo-cli