cc-audit: A Skill That Forces You to Save Claude Code Quota

Claude has been messing with quotas lately and everyone’s complaining. Complaining aside, Claude is still what I use to work. A while back I posted a long list of quota-saving tips on Threads, which ended up getting a thousand-plus likes. Someone suggested it should be a skill, so I built one.

Here it is: https://github.com/danyuchn/cc-audit

Hand the link to Claude Code and ask it to install.

What It Does

cc-audit is a health-check skill. When triggered, it runs four steps and hands back a structured report and an actionable fix list.

Step 1: Static Config Audit

• Is your default model set to Sonnet? Sonnet is much cheaper than Opus for most tasks and usually enough.
• What hooks are configured? Any runaway post-hooks silently eating tokens?
• MCP server list—each one is a fixed startup overhead.
• Line counts per rules file—catches rules that have bloated past 200 lines and need splitting.
• Skills list.

Step 2: JSONL Session Analysis (last 5 sessions)

This is the meat of it. It pulls your five most recent JSONL files and computes, per session:

• Context usage: (input_tokens + cache_creation + cache_read) / 200000
• Cache type: distinguishes Pro’s 5-minute cache (ephemeral_5m_input_tokens) from Team/Enterprise’s 1-hour cache (ephemeral_1h_input_tokens)
• Compact triggers: detects /compact events via type=system messages with a summary key
• Image-triggered cache misses: spots the pattern “user content has image type + next assistant message has cache_read < 500”

After the run you get a clear view per session: how many tokens it cost, whether auto-compact ran, whether a pasted image nuked the cache, whether context usage was 80% or 30%.

Step 3: Check Against Official Best Practices

This step WebFetches the latest Anthropic Claude Code docs and compares your current setup against official recommendations. Are the tips I wrote getting outdated? They can’t, because it re-fetches every run.

Step 4: Rules Path-Specificity Check

It checks whether your rules should have a paths: condition so they only load when relevant file paths show up. This one saves a lot, and most people don’t know about it.

Why Not Just Use /stats?

Claude Code’s built-in /stats tells you “this session used X tokens,” but it doesn’t tell you:

1. Why it used that much
2. Which settings are the culprits
3. What to change next

cc-audit’s output is an actionable fix list. Each item points to a specific file, specific line, specific change. You don’t have to cross-reference docs and guess.

Build the Habit

My own usage is to run cc-audit every week on Claude Max reset day. Pair this with the workflow I wrote about in Obsidian + Claude Code—the day your weekly quota is about to reset and you’re temporarily a “single-day token millionaire” is exactly when you should be doing this kind of refactor.

I used to do this by hand, cross-checking docs, and it took 30-40 minutes. As a skill, it’s fully automated in five minutes and I stop forgetting items.

Hand this link to Claude Code: https://github.com/danyuchn/cc-audit

It’ll install and verify, then you just type /cc-audit to run.