reflect

name	reflect
description	Analyze the current Claude Code session for efficiency — what it cost, where tokens were wasted, and what to add to CLAUDE.md to prevent it next time.
invocation	explicit

Run when the user types /reflect or asks to review session efficiency, analyze this session, or find what could have been done better.

What to do

1. Run the reflect analyzer:

   node "${CLAUDE_PLUGIN_ROOT}/index.js" --reflect --project "$PWD" 2>/dev/null || claude-trace --reflect --project "$PWD"
   

   Use the Bash tool. It finds the most recent session automatically.

2. Reason about the output:

   • Which tools dominated cost (>50% = attention needed)?
   • Are there retry loops (same tool called 3+ times in a row)?
   • What's the tool error rate?
   • Were there correction turns where the human redirected Claude?
   • Was the opening prompt specific enough to avoid back-and-forth?

3. Output 3–5 specific insights. Format each as:

   "Turns X–Y: [what happened]. Cost ~$X. [Why it was friction]. Fix: [one concrete suggestion]."

   Examples:

   • "Bash was called 6 times in a row (~$1.07 wasted). Batch commands into a script instead."
   • "Read was called 3x on the same file after edits. Add: 'read files upfront, not repeatedly' to CLAUDE.md."
   • "4 correction turns in 12 prompts. Opening prompt was underspecified — include file paths and expected output."

4. Offer CLAUDE.md additions. For each insight, propose the exact text. Ask: "Want me to append this to your CLAUDE.md?"

Cross-session analysis

For patterns across recent sessions:

node "${CLAUDE_PLUGIN_ROOT}/index.js" --reflect --project "$PWD" --recent 5 2>/dev/null || claude-trace --reflect --project "$PWD" --recent 5

Look for: recurring retry loops, growing costs over time, consistent error patterns.

What NOT to do

• Don't summarize session content — only analyze efficiency
• Don't report token counts without context on whether they were well-spent
• Don't suggest additions already present in CLAUDE.md