| name | cost-audit |
| description | Analyzes the current session's token usage, pinpoints where tokens were wasted, and produces concrete optimization actions |
| user-invocable | true |
| allowed-tools | Bash(claude*), Read, Grep |
| effort | low |
Goal
Make it visible where the tokens in this session actually went. Not guesses — evidence. Output: "do these 3 things next time and you save ~X%."
Step 1 — Capture current cost
/cost
Note the output: total tokens, input/output split, cache hit rate, total $.
Step 2 — Find the leaks
Scan the transcript for these signals:
| Signal | How to detect it | Why it wastes tokens |
|---|
| Unnecessary file reads | Read calls touching node_modules/, dist/, *.lock, coverage/ | Missing .claudeignore |
| Re-reading the same file | Same path appears in multiple Read calls | Context lost — /compact should have happened earlier |
| Reading entire large files | Read without limit on files >500 lines | Should have used offset/limit or Grep |
| Bash + grep/cat/find | Bash calls to grep, cat, find, ls -R | Dedicated tools (Grep/Glob/Read) are cheaper and more precise |
| Vague prompt → many tool calls | Single-word asks ("improve", "fix", "look") followed by 5+ tool calls | Prompt wasn't specific |
| Repeated onboarding research | Same concept re-discovered across sessions | Belongs in CLAUDE.md |
Step 3 — Report
Give the user a report in this shape:
## Session summary
- Total: <tokens> (<$>)
- Cache hit: <%>
## Top 3 leaks
1. [Evidence]: X files under src/components/ were read in full; ~Y tokens combined
Fix: add to .claudeignore → <lines>
Estimated savings: ~Z%
2. ...
3. ...
## Action list
- [ ] Create .claudeignore with <lines>
- [ ] Add a "<topic>" section to CLAUDE.md
- [ ] Use `/compact <hint>` earlier in long sessions
Step 4 — Optional auto-fixes
If the user approves:
- Create
.claudeignore if missing, add the detected paths
- Create a
CLAUDE.md scaffold if missing, seed it with the repeated-research topics
- Suggest a specific
/compact hint phrase for this project
Rules
- No guessing — only report what's in the transcript
- Always a number — not "a lot", but "12,400 tokens"
- Action-oriented — every finding ends with a concrete command or edit
- Keep it short — 3 top leaks is plenty; skip the long tail
