Research Agent
Search, Summarize, Cite
Why Research Is the Killer Feature
Email and calendar management save time. Research assistance saves *thinking*. An AI assistant that can search your saved knowledge, synthesize findings, and cite sources transforms how you make decisions.
Search: Finding What Matters
Your bookmarks are a personal knowledge base — articles, documentation, tools, and references you've saved. The search tool needs to be smarter than exact string matching.
A good bookmark search checks:
Fuzzy matching (case-insensitive, partial matches) makes the search useful for natural language queries like "that article about API rate limiting" even if the bookmark title is "Anthropic API Best Practices."
Summarization with Citations
Every claim your assistant makes should trace back to a source. This isn't just good practice — it's what makes AI output trustworthy.
The pattern:
[Source: bookmark title]Example output:
> Rate limiting best practices suggest using exponential backoff starting at 1 second [Source: Anthropic API Best Practices]. For high-throughput applications, implement request queuing with a token bucket algorithm [Source: Building Resilient API Clients].
Research Brief Format
A structured research brief is more useful than freeform prose:
| Section | Purpose |
|---|---|
| Topic Overview | 2-3 sentence summary of what was found |
| Key Findings | Bulleted list of facts, each with a citation |
| Source List | All bookmarks referenced, with URLs |
| Knowledge Gaps | What the available data doesn't cover |
| Recommended Actions | Suggested next steps based on findings |
This format works because it's scannable, actionable, and honest about limitations.
Cross-Referencing Research with Tasks
Your task list often contains items that need research:
By connecting the research agent to the task list, Claude can answer: "What research do I need for the infrastructure migration project?" It searches tasks for research-related items, then checks bookmarks for existing material on those topics.
Building Trust Through Transparency
| Practice | Impact |
|---|---|
| Always cite sources | User can verify claims |
| Flag knowledge gaps | User knows what's missing |
| Distinguish facts from inference | User knows what Claude concluded vs. what was stated |
| Show search results | User sees what data Claude used |
An assistant that says "I found 3 bookmarks on this topic, but none of them cover pricing" is more useful than one that guesses at pricing.
Key Takeaways
This is chapter 4 of Build Your AI Assistant with Claude.
Get the full hands-on course — free during early access. Build the complete system. Your projects become your portfolio.
View course details