lean-claude

lean-claude

Token and context efficiency for Claude Code — spend fewer tokens without losing quality.

Install

/plugin marketplace add merkle-ne-playground/lean-claude
/plugin install lean-claude@lean-claude

What's inside

• token-efficiency skill — loads the six core efficiency rules into Claude's working context when you ask it to "save tokens", "be lean", or "reduce context spend".
• /token-audit — read-only inspection of your Claude Code config and installed skills; flags oversized SKILL.md files and missing token-saving config, and outputs a prioritised list of recommended changes.
• /token-setup — interactive wizard that applies the recommended changes from /token-audit, with an explicit confirmation prompt before every write.

Usage

Ask Claude to work leanly in plain language — the skill triggers on phrases like "save tokens", "be lean", "reduce context":

> keep this session lean on tokens

Audit your setup, then apply the safe recommendations:

/token-audit      # read-only: reports oversized skills + config wins, writes nothing
/token-setup      # applies the audit's recommendations, confirming before each write

A typical /token-audit report flags skills whose SKILL.md exceeds 20 KB, suggests skillOverrides for skills you rarely auto-trigger, and points out a missing or oversized skillListingMaxDescChars. /token-setup then walks each change with a yes/no prompt — nothing is written without your confirmation, and existing config is merged, never overwritten.

Practices

Six rules that the token-efficiency skill enforces:

1. Subagent routing — pin each agent to the right model tier: Haiku for read-only/search tasks, Sonnet for edits and generation, Opus for architecture and deep reasoning.
2. Cache hygiene — don't /clear mid-task, keep turns under 5 minutes apart, append context rather than rewriting early turns, and keep tool-call order stable.
3. Narrow reads — grep before reading; read only the hit range with offset/limit; never re-read a file after an Edit.
4. /clear discipline — clear only between unrelated tasks; stale task context is re-billed on every turn.
5. Big-skill refactor — keep SKILL.md bodies under 20 KB; move detail into reference files that are read on demand, not on every invocation.
6. Skill-listing trim — use skillOverrides to set unused skills to name-only or off, and cap skillListingMaxDescChars to shrink the per-turn listing cost.

Complementary

caveman by JuliusBrussee (MIT) — an ultra-compressed communication mode that cuts roughly 75% of output tokens by dropping filler while keeping full technical accuracy. Pairs well with lean-claude. Third-party plugin; referenced here for discoverability, not bundled.

Contributing

Issues and PRs welcome. The plugin is intentionally small — keep additions lean:

• Skill content lives in skills/token-efficiency/. Keep SKILL.md thin (under ~4 KB); put depth in reference/ files that are read on demand. New guidance usually means a new or extended reference file, not a fatter core.
• Commands are prompt-type markdown in commands/ — no scripts, no runtime dependencies. /token-audit must stay read-only; /token-setup must confirm before every write and merge rather than replace config.
• No session hooks. A SessionStart hook that injects content every session would defeat the plugin's purpose. Don't add one.
• Run a quick install smoke test (/plugin marketplace add + /plugin install) before opening a PR.

License

MIT — see LICENSE.