session-reflection-analysis

name	session-reflection-analysis
description	Use when asked to reflect on how the session went

Session Reflection Analysis

Analyze recent chat history to identify improvement opportunities and reduce token waste in future sessions.

Overview

This skill helps identify patterns of inefficiency in Claude Code sessions by analyzing session history. The analysis focuses on actionable improvements to documentation, automation, and workflows.

Step 1: Locate Project Sessions

First, find the correct project folder. Claude Code stores sessions at ~/.claude/projects/ with folder names derived from the project path (slashes become dashes, colons become double-dashes).

# List available project folders
ls -la ~/.claude/projects/

# Find the current project's session folder by matching the working directory
# Example: /Users/harper/src/myproject -> -Users-harper-src-myproject
CURRENT_PATH=$(pwd | sed 's|^/||; s|/|-|g')
PROJECT_DIR="$HOME/.claude/projects/-${CURRENT_PATH}"

# Verify it exists
if [ -d "$PROJECT_DIR" ]; then
    echo "Found project dir: $PROJECT_DIR"
    ls -la "$PROJECT_DIR"/*.jsonl 2>/dev/null | head -5
else
    echo "No session folder found for current project"
    echo "Available projects:"
    ls ~/.claude/projects/
fi

Step 2: Generate Session Summary

CRITICAL: Do NOT read raw session files directly. They are massive and will consume your entire token budget.

First, verify jq is available:

if ! command -v jq &> /dev/null; then
    echo "ERROR: jq is required but not installed"
    echo "Install with: brew install jq (macOS) or apt install jq (Linux)"
    exit 1
fi

Then use jq to generate a 98% token-reduced summary:

# Set project dir (adjust if auto-detection didn't work)
PROJECT_DIR="$HOME/.claude/projects/-Users-harper-Public-src-2389-matrix-productivity"  # Update this path!
OUTPUT="/tmp/session-summary.jsonl"

# Optional: Only analyze sessions from last N days
DAYS_BACK=7
CUTOFF_DATE=$(date -v-${DAYS_BACK}d +%Y-%m-%d 2>/dev/null || date -d "${DAYS_BACK} days ago" +%Y-%m-%d)

# Summarize: extract user requests, tool names, assistant text (truncated)
cat "$PROJECT_DIR"/*.jsonl 2>/dev/null | jq -c '
select(.type == "user" or .type == "assistant") |
{
  type,
  ts: .timestamp,
  content: (
    if .message.content | type == "string" then
      .message.content[0:300]
    elif .message.content | type == "array" then
      [.message.content[] |
        if .type == "text" then {t: "text", v: .text[0:300]}
        elif .type == "tool_use" then {t: "tool", v: .name}
        elif .type == "tool_result" then {t: "result", len: (.content | length)}
        elif .type == "thinking" then empty
        else {t: .type}
        end
      ]
    else null
    end
  )
}' > "$OUTPUT" 2>/dev/null

echo "Summary: $(wc -l < "$OUTPUT") messages, $(wc -c < "$OUTPUT" | xargs) bytes"

This extracts:

• User message text (first 300 chars)
• Tool calls (name only, not full output)
• Tool result sizes (not content)
• Timestamps for sequence analysis

Step 3: Launch Subagent for Analysis

Use the Task tool to spawn an analysis subagent:

Task tool parameters:
- subagent_type: "Explore"
- prompt: |
    Analyze the session summary at /tmp/session-summary.jsonl for inefficiency patterns.

    Read the file and look for these patterns:

    | Pattern | Example | Impact |
    |---------|---------|--------|
    | **Wasted tokens** | Re-reading same file 5+ times | High |
    | **Wrong paths taken** | Implemented feature, then discovered existing code | Medium-High |
    | **Repeated mistakes** | Same error type in 3+ instances | Medium |
    | **Missing automation** | Manual steps repeated across sessions | Medium |
    | **Missing documentation** | Had to figure out what should be in CLAUDE.md | Medium |
    | **Unnecessary tool calls** | Called multiple tools when one would work | Low-Medium |
    | **Context loss after compaction** | Info that should survive in persistent docs | High |
    | **Assumption without verification** | Decisions made without checking existing code | High |

    For each pattern found, propose improvements as:
    1. CLAUDE.md updates
    2. New skills for .claude/skills/
    3. New slash commands for .claude/commands/
    4. Script automation
    5. Git hooks
    6. Workflow changes

    Make all proposals copy-paste ready with complete text/code.
    Quantify impact where possible (tokens saved, time saved).
    Be specific with examples from the actual session data.

Step 4: Create Reflection Document

The subagent should generate a document with this structure:

# Session Reflection: YYYY-MM-DD

## Summary Statistics
- Session date range: [start] to [end]
- Messages analyzed: [count]
- Major inefficiencies found: [count]

## Proposed Improvements

### High Priority

#### 1. [Problem Title]

**Problem observed**: [What went wrong - be specific with examples]

**Proposed solution**:

[Complete text/code to add or change - copy-paste ready]

**File to modify**: `path/to/file.md`

**Cost**: [Effort to implement, any downsides]

**Benefit**: [Time/tokens saved - quantify if possible]

---

### Medium Priority
...

### Low Priority
...

## Implementation Notes

[Cross-cutting concerns, dependencies, warnings]

Step 5: Present Findings

After analysis completes:

1. Summarize key findings - 3-5 bullet points of major inefficiencies
2. Show the reflection document path - Where to find full details
3. Ask which changes to implement - Don't auto-implement

Critical Rules

1. NEVER read raw session files directly - Always use the jq summary
2. NEVER implement changes automatically - Only propose them
3. Make proposals copy-paste ready - Complete text, not descriptions
4. Quantify impact when possible - "Saved 10K tokens" not "saves tokens"
5. Be specific with examples - "Re-read config.py 7 times" not "read files multiple times"

Example Analysis Output

High Priority Issues Found:

1. **Re-reading the same file repeatedly** (7 times in one session)
   - Propose: Add key architecture summaries to CLAUDE.md
   - Benefit: Save ~15K tokens per session

2. **Implemented feature without checking existing code first**
   - Took 3 iterations to match actual behavior
   - Propose: Add "Check existing code FIRST" rule to CLAUDE.md
   - Benefit: Prevent wrong-path implementations

3. **Manual repetitive commands**
   - Same shell commands typed each session
   - Propose: Create slash commands for common operations
   - Benefit: Save 2-3 minutes per session, prevent typos

Troubleshooting

Can't find project folder

# List all projects with recent activity
ls -lt ~/.claude/projects/ | head -10

# Search for a keyword in project names
ls ~/.claude/projects/ | grep -i "keyword"

No jsonl files found

Sessions may not have been saved yet, or the project path detection failed. Check the folder manually.

jq not installed

# macOS
brew install jq

# Ubuntu/Debian
sudo apt install jq

# Or use the standalone binary
curl -Lo /usr/local/bin/jq https://github.com/jqlang/jq/releases/download/jq-1.7.1/jq-macos-arm64
chmod +x /usr/local/bin/jq