usage-limit-reducer

Diagnose Claude Code token usage from local JSONL conversation logs.

1|Updated May 7, 2026

Search Tags:#claude code #prompt optimization #cost reduction #token usage #usage analytics #local jsonl

Ready to run

npx skills add https://github.com/Soham407/studio-kickstart --skill usage-limit-reducer

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: usage-limit-reducer
Source: https://github.com/Soham407/studio-kickstart/tree/main/architecture/usage-limit-reducer
Command: npx skills add https://github.com/Soham407/studio-kickstart --skill usage-limit-reducer

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires python3, argparse, json, datetime, collections, and includes scripts (resource) components.

What problem does it solve?

It helps you stop burning tokens in Claude Code by identifying where your spending actually goes (often re-reading long history) and guiding you to reduce usage with targeted fixes instead of generic advice.

Core Features & Use Cases

Measures real token/cost usage locally: Runs a usage report over Claude Code’s JSONL logs in ~/.claude/projects/*/*.jsonl to show token breakdowns by model, project, and day, including cache-hit rate and callouts.
Diagnoses your current session: Checks factors like conversation length, whether CLAUDE.md exists, and which model you’re using to determine which reduction rules are most relevant.
Applies only the most relevant rules safely: Recommends 2–4 actions (for example, /compact, /clear, switching to a cheaper model for simple edits, creating/updating CLAUDE.md) and acts only after confirmation, with no silent changes and no network data sharing.

Quick Start

Ask your Claude Code agent for: “Run usage-limit-reducer and tell me which 2–3 steps I should take to stop wasting tokens.”

Frequently Asked Questions about usage-limit-reducer

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I reduce Claude Code token usage from long conversations?▼

To reduce Claude Code token usage, you can analyze local JSONL conversation logs to diagnose re-reading patterns and apply targeted fixes like compacting history or clearing context. This approach identifies where your spending actually goes to cut token waste with real usage data.

What is the best way to track Claude Code token costs by model and project?▼

The best way to track Claude Code token costs is by running a local usage-report script over your JSONL logs in ~/.claude/projects. This generates a token breakdown by model, project, and day, including cache-hit rate, to show exactly where your spending goes without sharing network data.

Do I need Python to analyze local JSONL logs for token spend patterns?▼

Yes, you need Python3 to analyze local JSONL logs for token spend patterns, as the diagnostic script relies on standard Python libraries including argparse, json, datetime, and collections. These dependencies parse your local Claude Code conversation history to calculate usage metrics.

When should I use /compact or /clear to stop burning tokens in Claude Code?▼

You should use /compact or /clear during long chats or rapid token consumption when diagnostics indicate excessive history re-reading. A local log analysis recommends these actions only when they are the most relevant rules for your current session, ensuring you avoid unintended automatic changes.

Does switching to a cheaper model help with Claude Code cost reduction?▼

Switching to a cheaper model helps with Claude Code cost reduction when diagnostics identify model inefficiency for simple edits. By interpreting your model mix from local JSONL logs, you can determine if switching models is a relevant action to lower your overall token spend.

Can I diagnose Claude Code usage analytics without sharing my data externally?▼

You can diagnose Claude Code usage analytics entirely locally without sharing data externally. The analysis runs a bundled script over your local ~/.claude/projects/*/*.jsonl files, interpreting cache-hit and model mix to recommend next actions without any network data sharing.