Claude Code Usage Saver Kit
Find out where your Claude Code usage goes, then trim it.
Claude Code re-reads your whole conversation on every request, so long sessions, cache misses, idle loops and subagents use far more than your typing suggests. This kit gives you a 10-minute audit, twelve ranked habits taken from Anthropic's own documentation, optional lean settings, a hook that keeps test output out of your context, two small-model subagents, a context-budget skill and a read-only script that summarizes your own session logs.
What's included
- The 10-minute audit: read
/usage,/contextand/mcp, then match what you see to a likely cause and a first fix - Twelve habits, ranked by effort: clear between tasks, match the model, compact before breaks, rewind instead of compact, lower effort, trim MCP, specific prompts, plan mode, subagents, output filters, a short CLAUDE.md, no idle loops
- Lean settings: four optional
envlines (auto-compact window, goal check-ins, bash output cap, subagent model) with the trade-off for each, plus switches that depend on your habits - A test-output hook: adapted from Anthropic's cost guide so a green test run does not dump thousands of lines into your context, with the catch explained
- Two small-model subagents and a context-budget skill: read-only exploration, a noisy-command runner, and a skill that plans big tasks to read narrowly
- CLAUDE.md diet: a Compact instructions block and two prompts that shrink a bloated file and move workflows into skills
usage_audit.py: read-only summary of your local transcripts: longest sessions, volume from huge contexts, subagent share- Weekly routine, troubleshooting and a one-page cheat sheet
Try the audit script free
The same usage_audit.py that is in the kit, plus the cheat sheet and the Compact instructions block, is a free download. No signup.
Raw token counts explain volume, not the exact percentage on your plan meter. Use /usage for that.
Who it's for
Developers who use Claude Code on a subscription plan or an API key and keep hitting limits, or watching the bill, faster than their output suggests. It is most useful if you leave sessions open for hours, use subagents or loops, or have a long CLAUDE.md.
Format and honest notes
13-page PDF plus a ZIP of files. Written with AI assistance and checked against Anthropic's Claude Code documentation on 2026-10-17; the tool changes quickly, so the PDF links the pages to re-check. No file promises a specific saving. The audit script and hook were tested on sample data, not inside every Claude Code version. Not affiliated with Anthropic.
Questions
Will this make my Claude Code limits last longer?
It gives you a way to see where your usage goes and a set of habits and settings Anthropic's own docs recommend (shorter sessions, matching the model to the job, trimming output, smaller always-loaded context). It does not promise a specific saving: that depends on how you work, your plan and your model.
Is this affiliated with Anthropic?
No. It is an independent guide checked against Anthropic's public Claude Code documentation (the costs and environment-variable pages) on 2026-10-17. Claude Code changes quickly, so the PDF links the pages to re-check.
What exactly is in the ZIP?
A read-only Python audit script, lean settings lines with an explanation of each, a test-output hook, two small-model subagents, a context-budget skill, a Compact instructions block, prompts that shrink a bloated CLAUDE.md, and a one-page cheat sheet. Everything is a plain text file you can read first.
Does the audit script upload anything?
No. It reads the session transcripts Claude Code keeps on your machine and prints a summary. It never writes or uploads. The transcript format is not a public contract and can change, so the script says so if it finds nothing. We tested it on sample data, not on every Claude Code version.
Can I try something first?
Yes. The audit script and the cheat sheet are a free download on this page, and our free guide explains the main habits.
Want instruction files too?
The AI Coding Agent Config Kit ($14) has AGENTS.md, CLAUDE.md, Cursor and Copilot templates for 12 stacks. Together with this kit it is $17 instead of $21.