Usage Audit
Bloated context costs more and produces worse output. This skill finds the waste and tells you what to cut.
Step 1: Get /context Data
Check the conversation history for /context output. If the user already ran /context in this session, use that data. If not, ask:
"Run /context in this session terminal and let me know when you're done. I can't run slash commands myself, but once I can see the breakdown I'll audit everything it flags."
STOP HERE. Do NOT proceed to Step 2 until the user has ran /context. The context breakdown determines what to audit and in what order. Without it, the audit is guessing. Output the message above and wait for the user's next message.
Step 2: Audit What's Bloated
Based on the /context output, audit each category from largest to smallest. Run checks in parallel where possible.
MCP Servers
Each server loads full tool definitions into context every turn (~15,000-20,000 tokens each).
- Count configured servers from settings.json
- Flag any with CLI alternatives (Playwright, Google Workspace, GitHub all have CLIs that cost zero tokens when idle)
- Report total MCP overhead from /context output
CLAUDE.md
Read all CLAUDE.md files (project root,.claude/, ~/.claude/). Count lines. Then read every rule and test against five filters:
| Filter | Flag when... |
|---|---|
| Default | Claude already does this without being told ("write clean code", "handle errors") |
| Contradiction | Conflicts with another rule in same or different file |
| Redundancy | Repeats something already covered elsewhere |
| Bandaid | Added to fix one bad output, not improve outputs generally |
| Vague | Interpreted differently every time ("be natural", "use good tone") |
If total CLAUDE.md lines > 200, check for progressive disclosure opportunities: rules that only apply to specific tasks (API conventions, deployment steps, testing guidelines) should move to reference files with one-line pointers. Only recommend splitting when the file is actually bloated -- a lean CLAUDE.md with universal context is fine as a single file.
Skills
Scan.claude/skills/*/SKILL.md. For each skill:
- Count lines (flag > 200, critical > 500)
- Run the same five filters on instructions
- Check for restated goals, hedging ("you may want to"), synonymous instructions ("be concise" + "keep it short" + "don't be verbose")
Settings
Check settings.json for:
| Setting | Flag if | Recommended |
|---|---|---|
| autocompact_percentage_override | Missing or > 80 | 75 |
| BASH_MAX_OUTPUT_LENGTH (env) | At default (30-50K) | 150000 |
Hooks
Check settings.json for hooks configuration. For each hook:
- Flag hooks that run slow commands (network calls, heavy builds)
- Flag hooks with no timeout configured
- Check if any hook runs on high-frequency events (e.g., every tool call) where latency compounds
Memory Files
Glob ~/.claude/projects/*/memory/*.md. For each memory directory:
- Count total files and estimate total lines
- Flag stale memories (check if referenced project paths still exist)
- Flag memories > 50 lines (memories should be concise pointers, not docs)
- Flag memory directories with > 20 files (index bloat)
.claudeignore
Check if .claudeignore exists in the project root. If missing and the project has large generated/indexed directories, recommend creating one. Key patterns to check for:
- Large data directories (fixtures, seeds, dumps)
- Generated docs (typedoc, javadoc, storybook-static)
- Monorepo packages the user isn't actively working on
File Permissions
Check settings.json for permissions.deny rules. If missing, check whether bloat directories exist in the project:
| If this exists... | Should deny... |
|---|---|
| .git | .git |
| package.json | node_modules, dist, build,.next,.nuxt,.cache, coverage |
| Cargo.toml | target |
| go.mod | vendor |
| pyproject.toml / requirements.txt | pycache,.venv, *.egg-info |
Step 3: Score and Report
Score starts at 100. Deduct per issue:
| Issue | Points |
|---|---|
| CLAUDE.md > 200 lines | -10 |
| CLAUDE.md > 500 lines | -20 |
| Per 5 rules flagged by filters | -5 |
| Contradictions between files | -10 |
| Missing autocompact override | -10 |
| Missing bash output override | -5 |
| Skill > 200 lines | -5 each |
| Skill > 500 lines | -10 each |
| Per MCP server | -3 each |
| No deny rules + bloat dirs exist | -10 |
| Hook with no timeout on frequent event | -5 each |
| Stale or oversized memory files | -5 |
| Memory directory > 20 files | -5 |
| Missing.claudeignore + large dirs exist | -5 |
Floor at 0. Output this format:
# Usage Audit
Score: {N}/100 [{CLEAN|NEEDS WORK|BLOATED|CRITICAL}]
## Context Breakdown (from /context)
{Paste the key numbers from /context output}
## Issues Found
### [{CRITICAL|WARNING|INFO}] {Category}
{What's wrong}
Fix: {One-line actionable fix}
### Rules to Cut
{Each flagged rule: the text, which filter, one-line reason}
### Conflicts
{Contradictions between files, with paths}
## Top 3 Fixes
1. {Highest-impact fix}
2. {Second}
3. {Third}Score labels: 90-100 CLEAN, 70-89 NEEDS WORK, 50-69 BLOATED, 0-49 CRITICAL. Severity: CRITICAL > 10pts, WARNING 5-10pts, INFO < 5pts.
Step 4: Offer to Fix
After the report:
"Want me to fix any of these? I can:
- Show you a cleaned-up CLAUDE.md with the flagged rules removed
- Add the missing settings.json configs
- Add permissions.deny rules for build artifacts
- Add timeouts to hooks missing them
- Clean up stale memory files
- Generate a.claudeignore for this project
- Show which skills to compress"
Auto-apply settings.json and permissions.deny (safe, reversible). Show diffs for CLAUDE.md and skills -- let the user confirm before modifying instruction files.