← ClaudeAtlas

ccdash-analysislisted

Analyze a local ccdash archive of Claude Code and Codex requests — token spend, model routing, subagent fan-out, cost trend — and turn it into ranked, evidence-backed changes. Use when the user asks to analyze their ccdash usage, review token spend, work out where their Claude costs are going, find savings in model routing or subagent use, or explain a spend trend.
SeanoChang/seanock · ★ 0 · AI & Automation · score 70
Install: claude install-skill SeanoChang/seanock
# ccdash analysis ccdash keeps a local archive of every Claude Code and Codex request it has ingested: token counts and an equivalent cost, nothing else. This playbook turns that into advice. Read the archive through `ccdash export`. Answer the five questions in section 3, in that order. Cite a slice and a figure for every claim, and say "the data cannot answer that" rather than estimating — section 5 is the list of things it genuinely cannot answer, and it is not negotiable. ## 1. Get the data Run: ```bash ccdash export --stdout --format json ``` That prints one JSON document and writes nothing to disk. Add a window from what the user asked for; when they did not say, use `--range 30d`. | The user said | Flag | | ---------------------------- | ----------------- | | nothing about time | `--range 30d` | | today, so far today | `--range today` | | this week | `--range week` | | this month | `--range month` | | everything, all time, ever | `--range all` | Those five words are the only ones the flag takes. There is no `7d` and no `24h`; if the user asks for a window ccdash has no word for, take the next larger one and say in your answer that you widened it. Narrow further only when the ask is narrow: - `--tool claude` or `--tool codex` when the question is about one tool. - `--project /absolute/path` for one working directory. - `--session ID` for one session. - `--top N