research（研究）

Research

Core Philosophy

Research when getting it right matters. When current information saves hours of debugging, ensures secure implementations, or guides you to the right abstraction—research first.

Natural Triggers

Clear signals that research is needed:

• Hitting an error that smells like an API change
• Implementing something security-critical (auth, payments, file handling)
• Making architecture decisions you'll live with for months
• Working with libraries you know evolve rapidly
• That moment of "wait, is this still how we do this?"

Two Modes

Quick Check

When: Mid-flow verification Time: Under a minute Examples:

• "Is useEffect still the way to handle this in React 18?"
• "Did Stripe change their webhook payload?"
• "What's the current Node LTS version?"

Just search, grab the answer, keep coding. No storage, no ceremony, no permission needed.

Deep Dive

When: The decision really matters Time: 5-15 minutes Examples:

• Choosing between competing technologies
• Understanding a new architectural pattern
• Debugging something that doesn't match documentation

Always ask first: "This needs deeper research (5-15 min). Should I dig into this now?" Let the user decide if they want to pause for research or continue with existing knowledge.

Research thoroughly, save findings in research/[topic].md for team reference.

How to Research

1. Select Best Search Tool

Always use the best available web search. Priority order:

MCP servers (preferred when available):

• Tavily MCP server
• Exa MCP server
• Other specialized search MCP servers

Built-in tools (fallback):

• Cursor: web_search tool
• Claude Code: Built-in web search

Tell the user which you're using:

• "Using Tavily MCP server for enhanced search capabilities"
• "Using Exa MCP server for code-focused research"
• "Using built-in web search (no MCP servers configured)"

This transparency helps users understand tool selection and configure MCP servers if desired.

2. Search Strategy

Start with official sources - docs, changelogs, GitHub releases. Then expand to community discussions if needed.

3. Write Output

Output should be scannable and actionable. Skip the fluff, get to what matters.

Output Style

Do This

## Stripe Checkout v4 Migration

**Breaking change**: `redirectToCheckout()` removed in v4

**New pattern**:

- Use Payment Element (unified UI)
- Or Checkout Sessions API (hosted page)

**Migration**: [Specific code example]

**Source**: Stripe docs v2024.11.15

Not This

After extensive research into the evolving landscape of payment processing... It's
important to note that Stripe has made significant changes... Let's dive into the
migration process...

Real Examples

Quick Check in Action

[User struggling with Vite error]
Agent: "That error usually means a Vite config issue. Let me check v6 changes..."
[30 seconds later]
Agent: "Found it—Vite 6 requires explicit `server.fs.allow` for parent directories. Adding it now."

Deep Dive for Architecture

[User asking about state management for new project]
Agent: "State management in 2025 has some new players. Let me research current options for your scale..."
[10 minutes of research]
Agent: "For your use case (e-commerce, 50K users):
        - Zustand if you want simplicity (2KB, no boilerplate)
        - TanStack Store if you need framework agnostic
        - Redux Toolkit still solid for complex async flows

        Skip Redux unless you need time-travel debugging.
        Here's Zustand handling your cart state..."

Key Principles

Recognize patterns - When you see version-specific errors, deprecated methods, or post-2023 technologies, that signals research is needed.

Be transparent - Say "I should verify this" or "Let me check current best practices" rather than guessing.

Speed over perfection - For quick checks, first good answer wins. For deep dives, thoroughness matters.

No unnecessary storage - Quick research lives in the conversation. Only save deep research that others might reference.

Common Pitfall

Don't research everything. If your React knowledge from 2023 still works and the user isn't hitting issues, just build. Research is a tool, not a crutch.