/insights — Extract Session Insights for CLAUDE.md

Analyze Claude Code session transcripts to extract patterns that should be codified in CLAUDE.md files. Produces two-level output: Global entries (apply to all projects) and Project entries (apply to this project only).

Arguments

• current (default): Analyze the most recent session transcript
• latest: Analyze the 3 most recent session transcripts
• all: Analyze the 10 most recent session transcripts

Procedure

Step 1 — Locate Transcripts

Find session JSONL files for this project:

ls -lt ~/.claude/projects/*/

Identify the encoded project path that matches the current working directory. Sort by modification time (newest first).

Based on the argument:

• current / no argument: read the 1 most recent .jsonl file
• latest: read the 3 most recent .jsonl files
• all: read the 10 most recent .jsonl files

Step 2 — Extract Insights

Scan user messages in each transcript for five categories:

A. Corrections and Redirections Look for signals: “no”, “wrong”, “actually”, “instead”, “should be”, “not that”, “stop”, user restating the same instruction differently, or messages where the user interrupted and changed direction.

For each correction:

• What was Claude’s wrong assumption?
• What is the correct behavior the user wanted?

B. Coding Style Preferences Look for: naming conventions, formatting preferences, comment style, import ordering, file organization, language/framework choices.

C. Tool and Workflow Preferences Look for: preferred package managers, test runners, build tools, shell commands, editor workflows, git conventions, CI/CD preferences.

D. Communication Style Look for: preferred response length, level of explanation desired, when to ask vs. proceed, tone preferences, what annoys the user.

E. Architecture and Project Patterns Look for: module structure, dependency patterns, API design, error handling conventions, naming patterns specific to this project.

Step 3 — Categorize by Scope

Assign each insight to a scope:

Step 4 — Check for Duplicates

Read the existing CLAUDE.md files:

• ~/.claude/CLAUDE.md (global)
• ./CLAUDE.md (project, if it exists)

Compare each extracted insight against existing entries. Mark duplicates so they are not proposed again.

Step 5 — Present Findings

Output a findings table:

| # | Insight | Category | Scope | Duplicate? |
|---|---------|----------|-------|------------|
| 1 | ... | Correction | Project | No |
| 2 | ... | Coding style | Global | Yes (line 42) |

Then output draft CLAUDE.md entries organized by scope:

## Global Entries (for ~/.claude/CLAUDE.md)

- Entry 1
- Entry 2

## Project Entries (for ./CLAUDE.md)

- Entry 1
- Entry 2

Important

• Do NOT modify any files. Present findings for the user to review.
• The user decides what to add and where.
• If no meaningful insights are found, say so plainly rather than fabricating entries.
• Skip transcripts that are too short to contain useful patterns (fewer than 10 user messages).