Single tool 'deep_search' has a clear action verb and reasonable description (118 chars, within 10-1024 range). Input schema is visible with 5 parameters, all typed as string or integer. However, parameter descriptions are minimal (10-30 chars each), lacking guidance on constraints, defaults, or when to use each option. No output schema documented, LLMs cannot predict the markdown structure returned. Error handling returns generic text wrapping; no recovery guidance or categorization. Tool is read-only (safe), but lacks the depth expected for production use.

• Document the output schema: 'Returns a markdown string with aggregated content from top results, each prefixed with a citation link [source](url). Typical response is 2000-8000 chars.'
• Expand parameter descriptions with constraints and defaults: 'count: number of Brave results to fetch (default 5, max 20). Higher values increase latency. Use 10-20 for comprehensive searches, 5 for quick answers.' Similar detail for max_pages_to_fetch, max_chars_per_page, max_per_domain.
• Add error recovery guidance: catch specific errors (API rate limit, network timeout, invalid query) and return actionable messages like 'Brave API rate limited. Try again in 60 seconds or reduce count parameter.'
• Specify result limits and pagination: 'Returns aggregated markdown from up to max_pages_to_fetch results. Total response typically 2000-8000 chars. No pagination; for very large queries, reduce count or max_pages_to_fetch.'
• Add tool composition hints: 'Use this tool to gather background info before calling other tools. Results include citations; extract URLs for follow-up actions. Typical latency 3-8 seconds depending on count and max_pages_to_fetch.'