Added
- Dynamic research planning based on vague prompts via TodoWrite
- Structured research and question phases in evaluation workflow
- Support for 1-6 questions (increased from 1-2) to handle complex scenarios
- Explicit grounding requirement: questions based on research findings, not generic guesses
Changed
- Evaluation wrapper now creates custom research plans based on what needs clarification
- Research phase expanded to support any research method (codebase, web, docs, etc.)
- Removed prescriptive language about specific research tools
- Updated PROCEED criteria: "sufficient context" instead of "context from conversation"
- Token overhead increased to ~300 tokens (from ~250) due to enhanced instructions
- Final step clarified: "execute original user request" instead of "proceed with enriched prompt"
Improved
- More flexible and adaptive to different types of vague prompts
- Better grounding of clarifying questions in actual project context
- Clearer separation between research and questioning phases
- Numbered steps in Phase 1 and Phase 2 for better structure and clarity
- Preface moved to Phase 1 with context requirement explaining why clarification is needed
- Added specific examples for clarification reasons (ambiguous scope, missing context, unclear requirements)
- Critical rules repositioned under "ONLY ASK" section for better visibility during vague prompt evaluation
- Added "Do not rely on base knowledge" rule to prevent pattern-matching from training instead of research
- Step 2 clarified: "Research WHAT NEEDS CLARIFICATION, not just the project" with emphasis on online research for common approaches/best practices
- Step 3 simplified to "Execute research" (removed redundant warning)
- Step 4 explicitly requires using "research findings (not your training)" to prevent premature assumptions
- Specified recommended tools: Task/Explore for codebase, WebSearch for online research, Read/Grep as needed
Note: This release includes important improvements based on real-world usage feedback to prevent Claude from skipping research and relying on pattern-matching from training data.