Analyze Patterns

Analyze Patterns

When the user invokes /analyze-patterns or says "analyze sessions" / "find patterns":

Step 1: Extract messages from sessions

Run the extractor script, save to a temp file:

# Incremental (only new sessions since last run):
python3 ~/.claude/skills/analyze-patterns/scripts/pattern-detector.py --user-only --stats > /tmp/pattern-extract.json

# Or full rescan (all sessions):
python3 ~/.claude/skills/analyze-patterns/scripts/pattern-detector.py --full --user-only --stats > /tmp/pattern-extract.json

Use --full if the user says "all sessions" / "full" / "--full". Default is incremental. The --user-only flag extracts only user messages (smaller output).

Read /tmp/pattern-extract.json — it contains JSON with this structure:

{
  "session_count": 124,
  "projects": {
    "ProjectName": {
      "message_count": 25,
      "user_count": 25,
      "messages": [{"role": "user", "text": "..."}]
    }
  }
}

IMPORTANT: The file can be large (~400-500 KB). Read project by project, analyzing the largest ones first.

Step 2: Analyze patterns

Read the script output and find 4 types of patterns:

1. REPEATED_CORRECTION

User corrects Claude the same way in 2+ different sessions. Examples: "write shorter", "don't ask for confirmation", "use bun". → Suggestion: add to CLAUDE.md or create a rule file

2. FREQUENT_TOOL

The same workflow/skill/tool is used frequently. Examples: constantly calls /browser, always runs the same test command. → Suggestion: create an auto-trigger or shortcut

3. ABANDONED_TASK

Work started but not finished, errors left unfixed. Examples: config change started — session ended, bug found — not fixed. → Suggestion: create an issue to complete the work

4. CLAUDE_RESISTANCE

Claude didn't understand or the user had to explain again. Examples: Claude asks permission when told not to, writes in the wrong language. → Suggestion: improve the prompt/rule

Step 3: Show results to the user

For each pattern found, show:

• Type (REPEATEDCORRECTION / FREQUENTTOOL / ABANDONEDTASK / CLAUDERESISTANCE)
• Short title
• Evidence quotes from user messages
• Specific action suggestion
• Priority (high / medium / low)

Ask the user: "Create GitHub Issues for these patterns?"

Step 4: Create GitHub Issues (after confirmation)

For each confirmed pattern:

gh issue create \
  --repo / \
  --title "Pattern: " \
  --label "pattern" \
  --body "$(cat 

**Priority:** 
**Projects:** 

## Evidence

- "quote 1"
- "quote 2"

## Suggested Action

---
_Generated by /analyze-patterns on _
EOF
)"

Before creating, check for duplicates:

gh issue list --repo / --label pattern --state all --json title -q ".[].title"

Step 5: Update state

After a successful analysis:

python3 ~/.claude/skills/analyze-patterns/scripts/pattern-detector.py --full --update-state --stats > /dev/null

This marks all sessions as analyzed. The next incremental run will only pick up new ones.

Error handling

• No new sessions → "No new sessions since last analysis. Use --full for a full rescan."
• Empty script output → check that ~/.claude/projects/ exists and contains .jsonl files

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: denenis-lab
• Source: denenis-lab/my-claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.