/insights: consider stronger model or dedup against existing CLAUDE.md
Summary
The /insights command currently uses Haiku to analyze session transcripts and generate recommendations. While Haiku is cost-efficient for summarization, its limited reasoning leads to recommendations that duplicate rules/conventions already present in the user's CLAUDE.md files.
Problem
For users with well-maintained CLAUDE.md configurations, /insights surfaces suggestions that are already implemented — e.g., recommending conventions that are already documented in project instructions. This reduces the signal-to-noise ratio of the report and makes it feel less useful.
Root Cause
Haiku lacks the reasoning depth to cross-reference its recommendations against existing CLAUDE.md content. It generates suggestions based on session patterns without checking whether those patterns are already codified.
Suggestions
Any of these would improve the experience:
- Deduplication pass — Before finalizing the report, check recommendations against the user's CLAUDE.md files and filter out anything already covered
- Model flag — Allow
--modeloverride (e.g., Sonnet or Opus) for users willing to trade cost for quality - Smarter context — Feed existing CLAUDE.md content into the analysis prompt so the model can focus on genuinely novel insights
Expected Behavior
Recommendations in the report should only surface actionable items that aren't already present in the user's project configuration.
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗