[Claude Specialized] Methodology Guide (Long-form Reasoning/Claude Code Utilization)¶

This guide compiles practical guidelines based on Claude's (especially Sonnet/Opus) strengths: long-form reasoning, safety, and stable structured output. General-purpose techniques are consolidated in the ai-development side, while this page focuses specifically on "Claude-based" optimizations.

🔑 Usage Guidelines¶

Sonnet 4: Balance of stability and speed for long-form summarization/specification/multi-stage thinking
Opus 4.1: Highest precision for specification development/validation/complex problem decomposition
Haiku: Lightweight tasks/drafting/iterative assistance

🧩 Structured Output (Claude-oriented Types)¶

For strict YAML/JSON output, fix in order: "purpose → evaluation criteria → output schema → constraints"
Keep instructions short and staged, with minimal output examples (excessive examples induce hallucinations)

Example (generating specification framework):

Purpose: Generate specification draft for XXX
Evaluation: Consistency/implementability/testability
Output: YAML (sections: goals, scope, constraints, open_questions)
Constraints: Route unclear points to open_questions

🧱 Long-form & Segmented Reading Basics¶

Chunk strategy: Divide by chapters or responsibility units, provide headings and summaries first
Three-stage operation: "Summary → detailed questions → integration" to improve understanding stability

🛠 Claude Code Utilization Flow (Design → Implementation)¶

1) Specification (Sonnet/Opus): Problem decomposition and Definition of Done (DoD) definition 2) Implementation (Claude Code): Direct input of specifications into context for diff-driven development 3) Inspection: Specify review perspectives in bullet points (security/performance/readability) 4) Iteration: Promote failure cases to "prohibition list" and load at beginning of subsequent instructions

🧪 Avoiding Failure Patterns¶

Don't overload instructions (use staging)
Keep reference files to minimum set (reduce misreading)
Eliminate ambiguous terms (establish definition section every time)

For more general evaluation metrics/prompt structures, refer to ai-development practices.