Cost Optimizer

Audit a project's Claude Code setup for the handful of well-known cost leaks and output a prioritized remediation list. This is a static inspection — no pricing API calls, no guesswork about past spend.

When to use

• "My Claude Code bill doubled this month — what's going on?"
• "Audit this project for token waste"
• "Why is my session context always huge?"
• Post-mortem after a runaway agent run

What to check

Run each check and record PASS / WARN / FAIL with a short note.

1. CLAUDE.md size

wc -l CLAUDE.md .claude/CLAUDE.md 2>/dev/null

• PASS: under 100 lines
• WARN: 100-300 lines
• FAIL: >300 lines — this is prepended to every message, so every line costs tokens forever.

Fix: move reference material (API keys, runbooks, verbose examples) into skill files or docs/, and leave CLAUDE.md as a thin pointer.

2. Memory file bloat

find ~/.claude/projects -name 'MEMORY.md' -exec wc -l {} \;
find .claude/memory -type f -exec wc -l {} \; 2>/dev/null

• WARN if any memory file is >500 lines.
• FAIL if >1500 lines — memory should be a curated index with links to detail files, not a dump.

Fix: extract old sections into memory/<topic>.md detail files and link from the index.

3. Unused or shadowed skills

ls ~/.claude/skills/ .claude/skills/ 2>/dev/null

• Cross-reference with the skill invocations in the last 30 days of shell history or session logs if available.
• Skills that have never been invoked but have huge SKILL.md bodies (>300 lines) are loaded into the skill index every turn — prune them.

4. Subagent fan-out patterns

Grep the project for Task/agent launch patterns:

grep -rn 'TaskCreate\|subagent\|parallel.*agent' --include='*.md' --include='*.js' --include='*.ts' .

• WARN if the repo spawns >5 parallel subagents per run without aggregation — each spins up its own context window.
• FAIL if subagents re-read the whole repo instead of receiving a scoped prompt.

Fix: pass a short brief + a handful of file paths to each subagent, not the entire CLAUDE.md context.

5. Prompt cache hygiene

Look for .claude/settings.json and check if the project sets an unusually short cache TTL, or if hooks invalidate the cache every message (e.g. a hook that echoes a timestamp into the system prompt).

• FAIL any hook that injects date / uuidgen / random content into every turn — it kills the cache, 10x cost.

6. Model selection

Grep for explicit model pins:

grep -rn 'opus\|sonnet\|haiku\|claude-3\|claude-4' .claude/ CLAUDE.md 2>/dev/null

• WARN if Opus is pinned for routine tasks that Sonnet could handle. Recommend the /model-cost-compare skill for a per-task estimate.

7. Session file size

du -sh ~/.claude/projects/*/sessions 2>/dev/null | sort -h | tail -5

Huge session dirs suggest you never /compact long-running threads. Suggest periodic compaction.

Output format

Print a ranked list. Top of the list = biggest win per effort.

Cost Optimizer — <project>

Priority 1 (FAIL)
  [CLAUDE.md]      412 lines — prepended every turn. Extract sections into skills/.
  [cache hygiene]  .claude/settings.json hook injects `date +%s` on SessionStart. Remove it.

Priority 2 (WARN)
  [memory]         MEMORY.md is 820 lines. Index top, details linked.
  [model pinning]  Opus 4.6 pinned in 4 skill files — 2 could drop to Sonnet 4.6.

Priority 3 (nice-to-have)
  [skills]         3 skills (haiku-card, foo, bar) never invoked in last 30 days.

Estimated impact: roughly halving per-turn overhead — validate with a before/after /cost check.

Do not invent dollar savings. You can estimate relative impact ("~30% smaller context") but not absolute currency unless the user provides their current bill.

Example invocations

• /cost-optimizer
• "Audit this project for Claude Code cost waste"
• "My CLAUDE.md feels too big — what should I cut?"
• "Why is my prompt cache missing every turn?"