improve-session

name	improve-session
description	Analyze a Claude Code session and recommend improvements to reduce token waste, prevent linter loops, and automate repetitive workflows. Use when user says "improve session", "analyze session", "optimize session", or "/improve-session".
tools	Read, Write, Edit, Bash, Glob, Grep

Session Improver

Analyze Claude Code session transcripts and generate actionable recommendations to reduce token waste, prevent linter loops, and automate repetitive workflows.

Usage

• /improve-session — Analyze the most recent session
• /improve-session <session-id> — Analyze a specific past session
• /improve-session --current — Analyze the current live session

Workflow

Phase 1: Parse the session

Run the parser script to extract a structured summary:

ruby "${CLAUDE_SKILL_DIR}/scripts/parse-session.rb" <session-id-or-path>

If CLAUDE_SKILL_DIR is not set, fall back to the skill's location relative to this file. You can detect it by looking at the directory containing this SKILL.md.

The argument is either:

• A session UUID (found in ~/.claude/history.jsonl)
• --current for the most recent session
• A direct path to a .jsonl file

The script outputs JSON with these sections:

• linter_loops — Linter smells that triggered multiple edit cycles
• tool_failures — Tools that failed and were retried
• repeated_sequences — Workflow patterns that repeated 3+ times
• large_reads — Files read 3+ times in the session
• hook_failures — Hooks that failed repeatedly
• permission_events — Tools that needed human approval

Phase 2: Read context

For each finding, read relevant project files to understand what's already configured:

• The project's CLAUDE.md (if the session had a project path)
• The project's .claude/settings.json and .claude/settings.local.json
• The project's .claude/hooks/ directory
• Any .reek.yml, .rubocop.yml, eslint.config.*, etc.
• The global ~/.claude/CLAUDE.md
• The global ~/.claude/settings.json

Phase 3: Generate recommendations

Read the reference files for patterns and templates (relative to this SKILL.md):

• references/analyzers.md — What each finding type means
• references/fix-templates.md — Templates for each fix type
• references/ruby-linter-patterns.md — Ruby-specific patterns

For each finding, generate a specific recommendation:

1. Describe the problem — What happened, how many iterations, estimated token waste
2. Propose a fix — Concrete code/config to add
3. Specify placement — Where the fix should go:
   • project CLAUDE.md — Rules specific to this repo
   • global CLAUDE.md — Rules that apply everywhere
   • project settings.json — Hooks/permissions for this repo
   • global settings.json — Hooks/permissions everywhere
   • hookify rule — Behavioral guard
   • skill — Reusable workflow
4. Rank by impact — Most iterations/tokens saved first

Phase 4: Present and apply

Present findings in this format:

## Session Analysis: <session-id>

**Project:** <path>
**Duration:** X min | **Turns:** N | **Edits:** N | **Tool calls:** N

---

### Finding 1: <title> (N iterations)
**Impact:** ~Nk estimated wasted tokens
**Pattern:** <what happened>

**Recommendation:** Add to <location>:
```<config/code>```

**Apply this fix?**

Use AskUserQuestion to offer choices for each finding:

• Apply — Write the fix to the specified location
• Skip — Move to next finding
• Edit first — Show the fix, let user modify, then apply

When applying:

• For CLAUDE.md additions: append to the appropriate section, or create a new section
• For settings.json hooks: read the current file, merge the new hook config
• For hookify rules: create the .claude/hookify.<name>.local.md file
• For skills: create the .claude/skills/<name>/SKILL.md file stub

Important Notes

• Never remove existing configuration — only add to it
• If a fix already exists (e.g., the CLAUDE.md already has the rule), note it and skip
• For linter loop findings, check the project's linter config first — if it's already configured to ignore the smell, skip it
• Prefer CLAUDE.md rules over hooks for code style guidance (cheaper, no tool overhead)
• Prefer hooks over manual CLAUDE.md instructions for actions that should always run (formatting, linting)