Claude skills to reduce token usage (2026)

Google Autocomplete for "claude skills to" keeps offering the same cluster: reduce token usage, use less tokens, save tokens, optimize token usage, plus a GitHub variant. Those queries point at caveman. Intensity levels and the off-switch live in How to install and use the caveman skill. This page is the rest of the stack: shrink output, shrink files Claude reloads every session, measure the log, and keep a humanizer for copy a person will read.

Key takeaways

• caveman has 438,679 installs and 98,774 GitHub stars, with 92 installs in the last day (skills.sh registry, GitHub, August 18, 2026).
• Its SKILL.md targets roughly 75 percent fewer output tokens. The author's Honest Numbers page says that output figure is not a committed, reviewed benchmark, and the skill adds about 1 to 1.5k input tokens each turn.
• caveman-compress has 303,781 installs. It rewrites memory files such as CLAUDE.md. The repo reports about 46 percent input reduction across five fixtures.
• caveman-stats has 255,017 installs. It reads the Claude Code session log. The model does not estimate the number.
• When the deliverable is prose, use stop-slop (10,572 installs) or humanizer (4,310 installs). Caveman-speak is a cost control. It is a poor house style for docs.

Which Claude skills reduce token usage?

caveman leads this cluster by installs. It lives in Julius Brussee's GitHub repo (98,774 stars) inside a catalog of 59,924 skills and 147,337,564 recorded installs (Skillselion, August 18, 2026). The skill drops articles, filler, and pleasantries. Code, URLs, file paths, and error strings stay exact.

Two siblings in the same repo cover the rest of the bill. caveman-compress hits files that load every session. caveman-stats prints a number you can check against the provider usage page. Other GitHub repos ship a skill named caveman. Rank by installs and you land on juliusbrussee/caveman.

How do you install caveman and use less tokens?

From the project root:

npx skills add https://github.com/juliusbrussee/caveman --skill caveman

Trigger it with "caveman mode", "less tokens", "be brief", or /caveman. Default intensity is full. lite keeps sentences. ultra abbreviates. Install paths for any skill: How to install a Claude Code skill. Anthropic's skills docs explain where the files land on disk.

How do you cut input tokens with caveman-compress?

Output compression does nothing to CLAUDE.md. That file is paid on every turn. Run caveman-compress against memory files. It overwrites the original, writes a .original.md backup, and leaves code blocks, URLs, and paths untouched. The repo's fixture check averages about 46 percent fewer input tokens on five files. Treat that as a fixture average, not a forecast for a 4,000-line notes file.

npx skills add https://github.com/juliusbrussee/caveman --skill caveman-compress

Then /caveman-compress CLAUDE.md, or the path you actually load.

How do you know you saved tokens?

caveman-stats prints usage from the session JSONL. The Honest Numbers doc in the repo tells you to compare provider-billed totals on the same task with and without caveman. If the extra 1 to 1.5k input tokens per turn exceed the output you cut, turn it off. GitHub issue #145 is a terse coding Q&A that went net-negative. Issue #506 covers Copilot-style per-request billing: a shorter answer is still one request. Issue #550 is a Cursor A/B that spent more tokens with caveman on.

What Claude skills make it write like a human?

That Autocomplete line has its own page: Claude skills to write like a human. English drafts use stop-slop or humanizer. This page stays on token usage.

What else cuts token usage besides a skill?

Anthropic's session guide is the other half. /clear between unrelated tasks. /compact before a break so you summarize while the prompt cache is still warm. /context in a fresh session to see CLAUDE.md and MCP tool schemas before you type. Move procedures into skills, which load on demand. Disconnect MCP servers you are not using. Their tool schemas sit in every request. Official writeup: Maximizing the value of your Claude Code sessions.

If the agent feels slow because you stacked taste and planning skills, that search has its own page: Which Claude Code skills slow the agent down?.

A complete starter stack

• caveman for chatty sessions you will read yourself.
• caveman-compress on CLAUDE.md and other always-loaded notes.
• caveman-stats after a real task, then decide from the provider bill.
• stop-slop or humanizer when the output is a README, a guide, or anything a teammate will paste.

More agent skills: the AI agents shortlist and the AI agents category. Primer: What are Claude Code skills?.

Common pitfalls

• Treating the 75 percent SKILL.md target as a 75 percent smaller invoice. Output shrinks. Input and reasoning do not. The skill file itself costs tokens every turn.
• Leaving caveman on while writing docs, PR descriptions, or user-facing copy.
• Compressing CLAUDE.md and deleting the .original.md backup before you have read the result.
• Installing five "token optimizer" forks. The live catalog ranks juliusbrussee/caveman first by installs.

The GitHub repo is github.com/juliusbrussee/caveman. Read Honest Numbers before you trust a percentage.

Install counts come from the skills.sh registry via the Skillselion catalog, snapshot August 18, 2026.

Common questions