You hit your Claude Code limit on Thursday. Again.
And there was no way to tell which habit spent it. One free command reads the sessions already sitting on your disk and shows you where the week actually went: what you paid to carry context, how often the agent re-read a file it already had, and which tasks went to a model far bigger than they needed. It reads your transcripts, never your code.
Reads ~/.claude/projects. Prints to your terminal. Sends nothing.
Six answers you cannot currently get.
Not a score, and not a grade. Each one is a habit with a number attached, measured from what actually happened in your sessions rather than what you meant to do.
Where the money went
The share of your bill spent re-sending context rather than generating anything new. It is usually the largest single line item and usually invisible, because no invoice breaks it out.
Files you paid to read twice
How often the agent read a file it already had in context. Measured from content rather than activity, and reset by /compact — so it counts real waste, not how hard you worked.
Your own degradation point
The context size where your re-reads and tool errors start climbing. Not a generic number from a blog post — the knee fitted from your sessions.
What your config costs per turn
What your CLAUDE.md, skills, plugins, and MCP servers charge you on every single turn, before you have even stated the task. Sorted into dead weight, heavy-but-earning, and cheap-fine.
Tasks that went to too big a model
Which sessions were routed to a model several tiers above what the work needed, and what the right-sized one would have cost instead.
What each subagent spawn carries
The prefix tokens every spawn drags along, and what that adds up to across a month of them.
It never reads your source code.
That is the line, and it is the first thing worth stating. A tool that reads your dev sessions has to be boring about privacy, so here is exactly what it looks at.
- It never reads your source code. The analysis runs on token accounting, tool calls, and file paths. Not the contents of the files you edited. If you are working on an employer's codebase, nothing about that codebase is examined.
- It runs on your machine. It reads the session transcripts Claude Code already writes to ~/.claude/projects. There is no account, no key, no sign-up, and no daemon left running.
- Nothing is sent unless you ask. A bare run prints to your terminal and stops there. Three flags change that and only if you type them: --judge and --open send a de-identified summary, and fix sends your config file for review. Your prompt gists never leave your machine.
- Nothing is left behind. npx it once and it is gone. fix writes proposed patches under ./.cc-audit/ and never applies them itself.
The exact contents of each of those three flags, how long anything they send is kept, and how to have it deleted are written out on the CLI data policy page.
The first one is the whole thing.
Step one is the product. The other two are optional, cost you nothing to skip, and are the only things that send anything anywhere.
Run it
No install, no account, no key. It reads the session transcripts Claude Code already writes to ~/.claude/projects.
Right-size your modelsoptional
Optional, and the flag is the consent. Sends a task gist plus metadata — never your code — to the hosted judge, then prints which sessions were over-modelled and what the cheaper tier would have cost.
Turn it into patchesoptional
Optional. Writes reviewable patches under ./.cc-audit/ — local model-pin edits plus a hosted CLAUDE.md trim. It never applies them itself.
Running this across an engineering org is a different question.
One laptop's logs answer it for one person. Whether the AI coding tools an organization bought are paying off, and where they are generating cleanup instead, needs more than that. That is Promptster for Teams, and it is a paid product.
Want this without running it yourself?
The command is the product and it stays free. A hosted version that keeps the same numbers up to date between runs, instead of you remembering to re-run it, is the thing being built next.
One email when it ships. No sequence, no newsletter. If it hasn't shipped within a year, your address is deleted — see the CLI data policy.
Frequently asked questions
Is it actually free?
Yes. It is a free command with no account, no key, no card, and no seat count. There is nothing to sign up for and no trial to expire.Does it work if I use Codex or Cursor instead of Claude Code?
Not yet. It reads the session transcripts Claude Code writes to ~/.claude/projects, so today it is Claude Code only.Will this tell me how to stop hitting my weekly limit?
It tells you what spent it, which is the part you cannot currently see. Most of what it finds is habit rather than volume: context you carried and never used, files re-read after they were already in the window, and tasks handed to a model several tiers larger than the work needed.Does it send my code anywhere?
No. It never reads the contents of your files at all — the analysis runs on token accounting, tool calls, and file paths. A bare run sends nothing at all; the flags that do send something are named on this page and you have to type them.What is the difference between this and Promptster for Teams?
This shows one developer their own habits on their own machine. Promptster for Teams answers a different question for an engineering organization, across many engineers, and is a paid product. The free command does not require or lead to it.Do my sessions get used to build benchmarks for other companies?
No. A bare run transmits nothing, so there is nothing to pool. The flags that do send a summary send a de-identified one, and individual sessions are not turned into org-level comparisons for anyone else.