audit-agent-overhead

Most "Claude got dumber" / "I burned through Max in 19 minutes" complaints trace to invisible per-turn overhead, not the model. This skill audits both scopes (global ~/.claude/, project .claude/) against nine known waste patterns and proposes targeted fixes.

Source: @mnilax thread, 2026 — 90-day instrumented study of 430h / $1,340 of Claude Code usage. Productive tokens: 27%. Overhead: 73%. The patterns are his.

When to run

• "Why am I hitting limits so fast?"
• "Claude feels slower / dumber / more expensive than last month"
• "Audit my Claude Code setup"
• "Trim my CLAUDE.md"
• After installing 3+ new plugins/skills/MCPs.
• Quarterly maintenance.

The nine patterns

Full detail and per-pattern fix recipe in patterns.md. Summary:

Total addressable: ~73% of tokens. Realistic recovery: 27% → ~65% productive.

Process

1. Run the audit

bash "${CLAUDE_SKILL_DIR}/bin/audit.sh"

${CLAUDE_SKILL_DIR} resolves to this skill's directory regardless of how it was installed (plugin cache, project .claude/skills/, or source repo). The script inspects three scopes and prints a per-pattern table with measured values vs. thresholds. It does not modify anything.

• Global — ~/.claude/CLAUDE.md, ~/.claude/settings.json, ~/.claude/skills/
• Project — ./.claude/CLAUDE.md, ./.claude/settings.json, ./.claude/skills/
• Plugin scope (the big one) — every plugin enabled in ~/.claude/settings.json:.enabledPlugins, cross-referenced with ~/.claude/plugins/installed_plugins.json to find install paths. SKILL.md files are deduplicated by skill name (plugins often ship the same skill under .claude/, .opencode/, and claude-plugin/ runtime variants — only one is loaded at runtime).

2. Read the output, score each pattern

For each pattern flagged "OVER":

• Summarize what's bloated (the specific files / hook entries / plugin names involved).
• Quote the relevant fix recipe from patterns.md.
• Estimate token savings using the share column above.

3. Propose fixes — one pattern at a time

Do not batch-edit. For each flagged pattern, present:

1. The current state (what was found).
2. The proposed change (exact edit, exact file path).
3. The risk (what behavior might change — usually "none if the rule was dead").
4. Ask: apply, skip, or modify?

Apply only what the user approves. Re-run the audit after each pattern to confirm the metric moved.

4. Hard rules

• Never delete a file outright. Move to .claude/.audit-backup-<date>/ so it's recoverable.
• Never edit ~/.claude/CLAUDE.md without showing a diff first. That file is the user's personal voice; tightening ≠ rewriting.
• Never disable a plugin/skill/MCP without naming it and asking. "Audit found 7 unused skills" is not consent — list them.
• Treat the user's CLAUDE.md rules as load-bearing by default. A "redundant" rule may be a scar from a past incident. When unsure, ask.
• Don't recommend /clear aggressively. Pattern 2 says cap at ~20 messages and use /compact, not nuke-and-restart.

5. Re-audit and report

After fixes, re-run bin/audit.sh and report:

• Per-pattern before/after (CLAUDE.md word count, hook count, MCP count, skill count).
• Which patterns are now within threshold.
• What's left and why (some patterns are behavioral, not config — flag those for the user to practice).

Anti-patterns

• Don't trust averages over baselines. If you don't have 7+ days of logs, the script will skip the "avg input tokens / prompt" line. Don't fabricate it.
• Don't generalize the article's percentages to this user. They're benchmarks, not measurements. Re-run the script to get actual numbers for this setup.
• Don't re-introduce overhead by writing a long fix-doc. If your "fix recipe" is itself 500 tokens, you've made the problem worse.