optimize.md

June 6, 2026 ยท View on GitHub

Analyze this project's Claude Code cost efficiency and provide an optimization report.

Do not ask clarifying questions. Run every step below, collect the data, and produce the final report.

Steps

1. Check CLAUDE.md

Read CLAUDE.md in the project root. If it exists:

  • Count the total number of lines.
  • Flag if it exceeds 150 lines (the recommended ceiling for cost efficiency).
  • Check whether it contains large code blocks, verbose examples, or sections that could be collapsed or moved to separate files.
  • Note whether it uses a concise, structured format (headings, bullet points, tables) vs. prose paragraphs.

If it does not exist, record that as a finding -- projects without CLAUDE.md cause Claude to spend extra tokens exploring the codebase on every session.

2. Check .claudeignore

Check whether .claudeignore exists in the project root. If it exists:

  • Read its contents and count the number of rules.
  • Assess coverage: does it exclude common high-token directories and files? Check for these specifically:
    • node_modules/, dist/, build/, .next/, __pycache__/, .venv/, venv/
    • Lock files: package-lock.json, pnpm-lock.yaml, yarn.lock, poetry.lock, Pipfile.lock
    • Generated files: *.min.js, *.min.css, *.map, *.bundle.js
    • Large data: *.csv, *.parquet, *.sqlite, *.db
    • Media: *.png, *.jpg, *.gif, *.mp4, *.ico
  • List any obvious gaps (common expensive directories or files that are not excluded).

If it does not exist, record that as a significant finding.

3. Check .claude/settings.json

Check whether .claude/settings.json exists. If it exists:

  • Read it and note:
    • Which model is configured as default (if any).
    • Whether any permission allowlists or denylists are set.
    • Whether --max-turns or any cost-related flags are configured.
  • Flag if Opus is the default for a project where Sonnet or Haiku would suffice for most tasks.

If it does not exist, note that the project is using global defaults.

4. Check MCP Servers

Run claude mcp list in the project root using Bash. If the command is not available or fails, check .claude/settings.json and ~/.claude/settings.json for mcpServers configuration instead.

  • Count the number of MCP servers connected.
  • Each server adds roughly 500-2000 tokens of tool schema to every turn.
  • Flag if more than 3 servers are connected (high overhead).
  • Flag any servers that appear unused or redundant.

5. Check Conversation Patterns

Look for signs of common cost pitfalls:

  • Check if there are custom commands in .claude/commands/ that could reduce repetitive prompting.
  • Check if the project has large files (>500 lines) that Claude is likely to read in full when targeted reads would suffice.
  • Count the total number of files in the project (excluding ignored directories) to estimate codebase complexity.

Scoring

Calculate a letter grade based on these criteria. Start at 100 points and subtract:

FindingPenalty
No CLAUDE.md-20
CLAUDE.md over 150 lines-10
CLAUDE.md over 300 lines-20 (replaces the -10)
No .claudeignore-20
.claudeignore exists but missing 3+ common exclusions-10
No .claude/settings.json (using global defaults only)-5
Default model is Opus with no task-based switching configured-10
4+ MCP servers connected-10
6+ MCP servers connected-15 (replaces the -10)
No custom commands defined-5

Grade scale:

  • 90-100: A+ (Highly optimized)
  • 80-89: A (Well optimized)
  • 70-79: B (Good, room for improvement)
  • 60-69: C (Average, notable savings available)
  • 50-59: D (Below average, significant waste)
  • Below 50: F (Unoptimized, immediate action needed)

Cost Estimate

Estimate the per-session cost using these assumptions:

  • Average session: 40 turns
  • CLAUDE.md is loaded every turn (line count x ~1.5 tokens per line x 40 turns)
  • Each MCP server adds ~1000 tokens per turn
  • Base system prompt: ~2000 tokens per turn
  • Average output per turn: ~500 tokens
  • Use the pricing for the configured model (default to Opus 4.8 at $5/$25 per 1M input/output tokens if no model is set)

Show the math briefly, then give the final estimate.

Report Format

Present findings in exactly this format:

CLAUDE CODE COST OPTIMIZATION REPORT
=====================================
Project: {project name from directory or CLAUDE.md}
Date:    {today's date}

CONFIGURATION AUDIT
-------------------
CLAUDE.md:          {EXISTS / MISSING} ({n} lines)
.claudeignore:      {EXISTS / MISSING} ({n} rules)
settings.json:      {EXISTS / MISSING} (model: {model or "default"})
MCP servers:        {n} connected (~{n} extra tokens/turn)
Custom commands:    {n} defined

COST ESTIMATE
-------------
Model:              {model}
Est. input/turn:    ~{n} tokens
Est. output/turn:   ~{n} tokens
Est. session cost:  ~${n.nn} (40 turns)
Est. monthly cost:  ~${n.nn} (assuming 5 sessions/day, 22 workdays)

GRADE: {letter grade} ({score}/100)

ISSUES FOUND
------------
1. {issue description}
2. {issue description}
...

RECOMMENDATIONS (ranked by impact)
-----------------------------------
1. [{HIGH/MED/LOW}] {recommendation} -- saves ~{n}% per session
2. [{HIGH/MED/LOW}] {recommendation} -- saves ~{n}% per session
...

ESTIMATED SAVINGS IF ALL APPLIED: ~{n}% reduction (~${n.nn}/session -> ~${n.nn}/session)

Keep the report factual and specific. Do not pad with generic advice. Every recommendation must reference a concrete finding from the audit above.