Claude Usage Analyst

Overview

Use this skill to produce evidence-based usage explanations from local ccusage data. Separate observed numbers from interpretation, and explain quota burn in human terms.

Workflow

1. Verify ccusage is available:

   ccusage --version

   If missing, install or update with npm install -g ccusage@latest or run with npx ccusage@latest.

2. Run the bundled analyzer for the requested window:

   python3 /path/to/claude-usage-analyst/scripts/analyze_claude_usage.py \
     --since YYYY-MM-DD --until YYYY-MM-DD --timezone Asia/Shanghai

   Default --since/--until is today in the selected timezone. For historical comparison, set --since to an earlier date such as the first day of the month; otherwise rank/median fields only describe the single target day.

3. If the user asks about a specific model comparison, pass aliases:

   python3 scripts/analyze_claude_usage.py --model-a fable --model-b opus-4-8

4. Read references/explanation-guide.md when writing the final answer.

Evidence Rules

• Base numeric claims on ccusage output or the bundled analyzer output.
• State the scope: ccusage claude measures local Claude Code usage logs, including Claude Desktop's Claude Code sessions when those local logs exist. It is not a complete ordinary Claude.ai chat bill.
• Report dates with timezone.
• Explain cache clearly: cache read tokens are still usage/quota pressure even though the user did not type those words.
• Do not infer Anthropic plan quota rules from local token counts unless the user provides plan details. Say "quota-like pressure" or "ccusage estimated cost/token burn" when exact plan accounting is unknown.
• When comparing models, compare both token volume and estimated cost. A model can have similar token volume but higher cost.

Output Shape

Use this structure unless the user asks otherwise:

1. Short conclusion in plain language.
2. Evidence table: total tokens, cost, input, output, cache create, cache read.
3. Model comparison table.
4. 5-hour block table when quota exhaustion is discussed.
5. Explanation of why the burn happened.
6. Confidence and caveats.

Keep the answer readable for non-technical users. Avoid unexplained terms like "cache read" without a one-sentence translation.