Research Docs

Research Docs

Overview

Your training data is outdated. Current documentation is always more accurate.

When designing features, planning implementation, writing code, or debugging issues involving libraries/frameworks/tools, you MUST fetch current documentation using Context7 before proceeding.

Core Principle

LLM training data becomes stale the moment training ends. Libraries evolve:

• APIs change between versions
• Best practices get updated
• New features get added
• Old patterns get deprecated

Never design or implement from memory. Always verify with current docs.

Mandatory Workflow

Step 1: Recognize the Trigger

You MUST use documentation search when you encounter ANY of these:

• Library name mentioned (react-query, fastapi, pydantic, express, etc.)
• Framework name mentioned (Next.js, Django, React, Vue, etc.)
• Version number specified (react-query v5, Python 3.12, etc.)
• Technical concept tied to specific tool (optimistic updates in react-query)
• Implementation questions (how do I X in Y?)
• Best practices questions (what's the right way to X?)
• Debugging library-specific behavior
• Evaluating technology choices for a design
• Designing a system that uses external libraries
• Any delta phase (design, plan, implement) involving libraries

Red flags that mean you're about to fail:

• "Based on my knowledge of..."
• "From what I remember about..."
• "The typical pattern for..."
• Writing code without checking docs first
• Designing around a library without checking current docs
• Assuming library capabilities from training data during design
• Having uncertainty about the correct approach

Step 2: Dispatch Documentation Search Subagent

Why subagent instead of direct Context7:

• Saves 10,000-20,000 tokens of context in main agent
• Subagent filters docs to only what you need
• Main agent stays focused on the task
• Better token management across the session

How to dispatch:

Dispatch the katachi:doc-researcher agent with the following information:

• Library name: Exact package/library name (e.g., "react-query", "fastapi", "pydantic")
• Topic: Specific concept or feature (e.g., "optimistic updates", "path parameters", "field validators")
• What you need: Specific APIs, patterns, or examples you're looking for

The agent will search Context7 documentation and provide a focused synthesis with:

• Exact API signatures
• Recommended patterns and best practices
• Code examples
• Version-specific guidance

Step 3: Act on Verified Patterns

After receiving subagent synthesis:

1. Cite what you learned: "According to react-query v5 docs (from doc-researcher)..."
2. Use exact API signatures provided
3. Follow recommended patterns from synthesis
4. Note any differences from what you expected
5. If gaps exist, dispatch another search or use WebSearch

During design phases:

• Base design decisions on current library capabilities
• Document technology choices with sources from synthesis
• Don't propose deprecated patterns as design options
• Verify feasibility of design approach with current API

During implementation:

• Use exact API signatures from synthesis
• Follow current best practices, not training data patterns
• Reference documentation source in code comments when the choice would be unclear

Never:

• Mix training data patterns with doc patterns
• Assume API names/signatures
• Skip documentation check "to save time"
• Design or implement first, verify later
• Use Context7 MCP tools directly (always dispatch katachi:doc-researcher agent)

Red Flags - STOP

If you're thinking ANY of these, you're about to violate the skill:

Context Rationalization Flags

• "I'm only using X% of budget" - Percentage hides absolute waste
• "Well within acceptable limits" - Ignores session-wide compounding
• "I have plenty of budget left" - Context is for ENTIRE session
• "This is just one search" - "Just one" becomes "just one more"

Efficiency Framing Flags

• "Direct access is more efficient" - You're optimizing for wrong metric
• "Subagent dispatch is overhead" - It's an investment, not overhead
• "Completed in fewer messages" - Messages don't matter, tokens do
• "For straightforward lookups, direct is optimal" - Context math doesn't change

Quality Justification Flags

• "I got comprehensive examples" - You don't need comprehensive, you need relevant
• "I can filter the docs myself" - Filtering doesn't remove docs from context
• "I need detailed information" - Subagent provides exactly what you need

The context math:

• Direct Context7: 15,000-25,000 tokens per search
• Subagent: 2,000-5,000 tokens per search
• Difference: 10,000-20,000 tokens SAVED per search
• 3 searches: 48,000 tokens saved
• That's 48,000 tokens for MORE searches, longer conversations, complex implementations

Never use "I have budget left" to justify waste.

When NOT to Use Documentation Search

Skip documentation search for:

• Trivial language built-ins (Python dict.get, JavaScript Array.map, string methods)
• Pure algorithms (sorting, searching, graph traversal)
• Questions about YOUR codebase (use Read/Grep)

But DO use documentation search for:

• Third-party libraries, even familiar ones
• Framework-specific patterns
• Version-specific APIs
• Best practices for tools
• Complex standard library modules (subprocess, streams, pathlib, logging)
• During design phases, when evaluating library fitness
• When comparing approaches that involve external libraries
• Even if the library was used recently in this session

When in doubt: dispatch a subagent. The cost of a subagent search (2,000-5,000 tokens) is trivially small. The cost of designing or implementing against stale docs is enormous.

Context Management Strategy

Why subagents are mandatory:

Context savings per search:

• Direct Context7: 15,000-25,000 tokens per search
• Subagent approach: 2,000-5,000 tokens per search
• Savings: 10,000-20,000 tokens per search

Across a session:

• 3 direct searches: ~60,000 tokens
• 3 subagent searches: ~12,000 tokens
• Savings: ~48,000 tokens

That's 48,000 tokens available for:

• More codebase files
• Longer conversations
• Additional library searches
• Complex implementations

Verification Checklist

Before claiming you've designed or implemented something correctly, verify:

• Dispatched katachi:doc-researcher agent to fetch current documentation
• Provided clear library name, topic, and what you need
• Received synthesis with API signatures
• API signatures match documentation exactly
• Patterns follow current best practices from synthesis
• No uncertainties remain about correct approach
• Can cite documentation source for key decisions
• Did NOT use Context7 MCP tools directly

If you have ANY uncertainty after receiving synthesis:

• Dispatch another katachi:doc-researcher agent with refined topic
• Use WebSearch for supplementary info
• Ask human for clarification

Never:

• Use Context7 MCP tools directly
• Ship uncertain design or implementation
• Skip documentation search to "save time"

Common Mistakes

Mistake 1: "I remember this API"

❌ "I know react-query uses useQuery, let me write this..."
✅ "Let me dispatch katachi:doc-researcher to verify the current useQuery API..."

Why it fails: APIs change. Your memory is from training cutoff.

Mistake 2: "Subagent overhead isn't worth it"

❌ "This is just one search, I'll use Context7 directly..."
✅ "Even one search saves 15,000 tokens. Always dispatch katachi:doc-researcher."

Why it fails: "Just one" becomes "just one more" throughout the session. Context compounds.

Mistake 3: "I'll verify after writing"

❌ [Writes full implementation] "Let me check if this is right..."
✅ [Dispatches katachi:doc-researcher first] "Now I'll implement using verified patterns..."

Why it fails: Fixing wrong code takes longer than writing correct code once.

Mistake 4: "This is just a design, I'll check docs during implementation"

❌ [Designs feature assuming library behavior] "I'll verify the APIs when I implement..."
✅ [Dispatches katachi:doc-researcher during design] "Let me verify this is feasible with current APIs..."

Why it fails: Wrong design foundations compound through the entire feature lifecycle. A design based on deprecated APIs wastes the spec, design, plan, AND implementation phases.

Summary

Before designing or implementing ANYTHING involving a library/framework:

1. Recognize trigger (library name → stop)
2. Dispatch katachi:doc-researcher agent
3. Provide clear library name, topic, and what you need
4. Receive synthesis with API signatures and patterns
5. Act using verified patterns from synthesis
6. Cite documentation source

Critical rules:

• NEVER use Context7 MCP tools directly
• ALWAYS dispatch katachi:doc-researcher agent for documentation
• Context savings: 10,000-20,000 tokens per search
• Your training data is always outdated
• Current documentation is always more accurate
• Dispatch agent first, design or write code second

This is not optional. This is mandatory.