tavily-search

Overview of tavily-search skill

What tavily-search does

tavily-search is a web research skill that uses the Tavily CLI to return search results in a format that works well for AI agents: snippets, relevance signals, metadata, and optional page content. It is best for users who need current information, source discovery, news checks, or a fast first step before deeper extraction and analysis.

Who should install tavily-search

The best fit for the tavily-search skill is anyone doing web research inside an agent workflow, especially when a normal model prompt lacks fresh data or trustworthy sources. It is useful for analysts, developers, content teams, and operators who need “search for,” “find recent coverage,” “look up sources,” or “what’s the latest on” behavior without building a search wrapper from scratch.

Why choose it over a generic web prompt

The main differentiator is structure. Instead of asking a model to vaguely browse, tavily-search calls tvly search directly and returns machine-friendly results that are easier to rank, filter, and chain into later steps. It also supports practical controls like domain filters, time ranges, topic selection, and search depth, which matter when you need current, narrow, or higher-recall results.

Important limits before adoption

This skill is only as usable as your Tavily CLI setup. If tvly is not installed and authenticated, the skill will fail. It is also a search-step skill, not a full research pipeline by itself: use it to discover sources and recent results, then extract, crawl, or synthesize afterward if the task needs more than snippets.

How to Use tavily-search skill

Install context and first-run setup

The tavily-search install path starts with the Tavily CLI, because the skill expects tvly on your PATH. The repository guidance is explicit: install and authenticate first, then run searches.

curl -fsSL https://cli.tavily.com/install.sh | bash && tvly login

If you are evaluating the repo, read skills/tavily-search/SKILL.md first. For broader CLI setup and auth alternatives, also check skills/tavily-cli/SKILL.md in the same repository.

How tavily-search is called in practice

In real usage, the tavily-search skill maps user intent into a tvly search ... --json command. Common patterns include:

tvly search "your query" --json
tvly search "quantum computing" --depth advanced --max-results 10 --json
tvly search "AI news" --time-range week --topic news --json
tvly search "SEC filings" --include-domains sec.gov,reuters.com --json

The most important inputs are:

• a precise query
• whether freshness matters
• whether certain domains should be included
• how many results you want
• whether broader recall (--depth advanced) is worth the extra noise

Turn a rough goal into a strong tavily-search prompt

Weak goal: “Find stuff about AI chips.”

Stronger request for tavily-search usage:

• “Search for recent reporting on AI chip export restrictions, prioritize Reuters, U.S. government sources, and major financial press, last 30 days, return 8 results.”
• “Find beginner-friendly articles explaining React hooks from official docs and trusted tutorials.”
• “Look up SEC filings related to Nvidia supplier risk, include sec.gov, recent results only.”

Why this works: the skill performs better when the request includes topic scope, freshness, preferred domains, and result count. Without those, the model may issue a broad search that is harder to use downstream.

Suggested workflow and output-quality tips

A practical tavily-search guide is:

1. Search broadly enough to discover source types.
2. Rerun with tighter filters.
3. Use the JSON output to compare relevance and source quality.
4. Only then move into extraction, crawling, or synthesis.

Quality tips that change outcomes:

• Use --topic news when the user clearly wants recent developments.
• Use --time-range when stale results would mislead.
• Use --include-domains when trust matters more than breadth.
• Use --depth advanced for harder research questions, but expect more cleanup.

tavily-search skill FAQ

Is tavily-search for Web Research worth installing?

Yes, if your agent frequently needs live source discovery. tavily-search for Web Research is especially useful when the task starts with unknown URLs and current information matters. If your work is mostly static internal docs or known websites, you may not need this skill first.

How is tavily-search different from asking a model to browse?

The tavily-search skill gives a repeatable command path and structured search output. That usually means less guesswork, better source filtering, and easier chaining into later steps. A generic prompt may work for casual browsing, but it is less reliable when you need explicit freshness, domain controls, or JSON results.

Is tavily-search beginner-friendly?

Mostly yes. The commands are simple, but beginners can get blocked on CLI install and login. If you want the fastest path, confirm tvly works in terminal before testing the skill. Once setup is done, the main learning curve is writing specific search instructions rather than vague topics.

When should I not use tavily-search?

Skip tavily-search when you already have the exact URL, need full-site crawling, or are doing non-web local analysis. It is also the wrong choice if your environment cannot install or authenticate the Tavily CLI.

How to Improve tavily-search skill

Give tavily-search better search intent

The biggest upgrade is input quality. Don’t ask for “info about X” when you really mean:

• recent news
• authoritative sources
• beginner explainers
• regulatory documents
• company coverage
• a fixed time window

A better tavily-search guide starts with the user’s actual decision need, not just the topic keyword.

Fix common failure modes

Typical problems are overly broad queries, no freshness constraint, too many low-value domains, and asking search to do full analysis. If results feel noisy:

• narrow the query
• add --include-domains
• set --time-range
• reduce or raise --max-results based on recall needs
• switch to --depth advanced only when basic search misses obvious sources

Iterate after the first result set

Good tavily-search usage is often two-pass. First pass: discover vocabulary, source patterns, and date ranges. Second pass: refine around the best terms and domains you found. This is usually better than trying to craft the perfect query upfront.

Improve the skill in the repository

If you want to improve the tavily-search skill itself, the highest-value additions would be clearer parameter examples for common research tasks, a quick decision table for --depth and --topic, and a few “bad query vs better query” examples. Since the skill currently lives mostly in SKILL.md, stronger examples would reduce adoption friction more than extra prose.