A Claude Code skill that optimizes Claude Sonnet 5 for lower token usage and lower cost.
Sonnet 5 is powerful but token-hungry by default: it over-explores codebases, rereads files it already has, spawns subagents/workflows for tasks a single tool call could solve, and gold-plates simple requests with unrequested abstractions and extra verification passes. lean-sonnet is a standing default-on skill that fixes this — it makes Claude Sonnet 5 pick the cheapest correct path on every task instead of the most expensive one.
- Token waste — rereading files/context already in hand, reading whole files instead of matched ranges
- Tool-call sprawl — spawning an Agent or Workflow for something one Read/Edit/Grep call solves
- Redundant verification — re-running the same check after it already passed
- Scope creep / gold-plating — unrequested abstractions, config flags, fallbacks, docs, comments
- Inefficient delegation — serial agent calls that should be batched, or top-tier models used for mechanical sub-tasks that a cheaper tier could handle
claude sonnet 5 optimization, reduce claude token usage, claude code cost optimization, anthropic sonnet 5 skill, claude agent skill, lower claude api cost, claude code token efficiency, claude sonnet 5 agentic workflow, cheaper claude code, claude 5 cost reduction, claude code skills directory
Drop SKILL.md into your Claude Code skills directory:
git clone https://github.com/<your-username>/lean-sonnet.git ~/.claude/skills/lean-sonnetClaude Code auto-discovers it from ~/.claude/skills/. For project-local use, clone into .claude/skills/lean-sonnet inside your repo instead.
To make it apply on every turn rather than relying on description-matching, add a pointer to it in your global ~/.claude/CLAUDE.md:
Apply the `lean-sonnet` skill's rules on every task: cheapest tool that
reaches "done", no unnecessary subagent/workflow fan-out, no rereading
context already in hand, one verification pass, no unrequested scope.MIT