Documentation-Driven Development Specification

When generating code or fixing bugs, consult official documentation and examples first before starting implementation.

⚠️ Core Mandatory Requirements

Scenarios Where Documentation Must Be Consulted

• When implementing new features, connecting to third-party library APIs, or troubleshooting API errors, consult official documentation first
• When encountering

  AttributeError

  /

  TypeError

  / uncertain version changes, must verify API definitions and examples
• For internal APIs encapsulated by the project itself, you can skip as appropriate

Agent Execution Steps

1. Stop guessing: Clarify the name and version of the API/module to be queried
2. Consult documentation: Prioritize using

   Context7

   to obtain official documentation; if unavailable, use

   DuckDuckGo

   to search for official sources and conduct secondary verification
3. Implement based on documentation: Implement or fix code according to the parameters, return values and examples in the documentation

MCP Invocation Baseline (Documentation Query Scenarios)

• For documentation/API issues, prioritize using

  Context7

  , and use

  DuckDuckGo

  for latest announcements or entry information
• Call at most one external service per round; if multiple tools are really needed, call them serially and explain the reasons
• Keep queries as minimal and necessary as possible (keywords, number of results, time window, tokens)
• When 429/rate limiting occurs, back off for 20 seconds and narrow the scope before retrying
• Attach "Tool Call Briefing" in the output (tool, input summary, parameters, time, source, retry)
• Offline first by default, do not upload sensitive information, comply with robots/ToS and privacy constraints

AI Agent Behavior Requirements

Problem Identification

AttributeError: 'X' has no attribute 'Y'

TypeError: unexpected keyword argument

Collect Error Information

• Complete error stack
• Relevant code snippets
• Version of the library/framework used

Implement Based on Documentation (Guessing is Prohibited)

# 错误：基于猜测实现
index.query(question, top_k=10)  # 可能参数名不对

# 正确：基于文档实现
# 查阅文档后确认正确的参数名
index.query(question, similarity_top_k=10)

Verification After Fix

• Run tests to confirm the fix
• Check if new problems are introduced
• Verify performance in different scenarios

References

• references/api-verification.md

  - Detailed description of API verification process (problem identification, documentation consultation, implementation of fixes)

• references/search-strategy.md

  - Documentation search strategy and question guidance

• references/mcp-usage-rules.md

  - MCP invocation rules (tool selection, rate limiting, security boundary, failure degradation and traceable output)