Dynamic Web Search Filtering

Instructions

Use this skill when you need to answer a question that requires searching the web and retrieving pages, but long HTML pages or irrelevant sections would otherwise waste tokens and degrade answer quality.

Workflow

1. Plan the search

   • Restate the user’s question as a small set of concrete sub-questions.
   • Decide what would count as sufficient evidence (exact values, quotes, definitions, etc.).

2. Search broadly, then narrow

   • Use web_search to collect several candidate sources.
   • Prefer primary sources (official docs, vendor pages, standards bodies) when possible.

3. Fetch only what you need

   • Use web_fetch on the most promising results.
   • If the fetched page is long, identify the sections likely to contain the answer.

4. Dynamically filter before reasoning

   • If the tool supports it, allow the tool to run code to post-process the retrieved content before it is loaded into context.
   • Otherwise, manually extract: headings + the smallest set of paragraphs/tables required to answer.
   • Discard navigation, unrelated sections, repeated footers, and boilerplate.

5. Answer with verifiable support

   • Provide the final answer using short quotes or precise figures.
   • Include citations to the fetched URLs.
   • If sources disagree, present both and explain the discrepancy.

Notes on cost and evaluation

• Web search is token-intensive; filtering reduces irrelevant context.
• Token costs can vary depending on how much code is written to filter context; evaluate against representative queries.

Examples

Example 1: API request enabling the newest web tools

See templates/api-request.json.

Example 2: Research task that needs careful filtering

User prompt:

“Find the current prices of AAPL and GOOGL, then calculate which has a better P/E ratio.”

Approach:

• Search for authoritative sources (major exchanges, financial data providers).
• Fetch the pages.
• Filter to the exact price and P/E fields.
• Calculate the comparison and cite the sources.

References

• Post: https://claude.com/blog/improved-web-search-with-dynamic-filtering
• Benchmarks mentioned: BrowseComp and DeepsearchQA (see post for details)