caveman-learn (Julius Brussee): a Claude skill that cuts your agent token costs, one approved edit at a time

caveman-learn acts on Caveman's token report: it proposes cost-cutting edits one at a time and applies each only after your yes.

Installs: 113.7K; on skills.sh
Stars: 116.7K; on GitHub
Files: 5; besides SKILL.md
Updated: Nov 2026; last repo commit

caveman-learn is the editing half of the caveman project's learn command. The caveman learn command measures where your agent's tokens go. This skill reads that report, shows the ranked token sinks, and applies fixes with your approval for each edit.

It runs the caveman command line tool, so install the CLI first (see the install box). It can trim a heavy CLAUDE.md, move re-pasted context into cavemem, and clean up duplicate or broken memory and rules files. After each edit it re-measures and reverts if tokens per turn did not drop.

Good for

Claude Code users with a heavy CLAUDE.md or repeated pasted context
Developers who want to see where an agent's tokens go and fix it
Anyone who wants savings measured, not guessed

Install caveman-learn

Claude Code

Copy this prompt.

Install the caveman-learn skill in Claude Code. Run this command:

npx @caveman-ai/cli@latest

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Requires Node.js 22.13 or newer.

caveman-learn runs caveman learn and caveman mem commands, so the caveman CLI must be on your machine. The README's one-command setup finds your agents and shows what it changes before you press Enter. INSTALL.md lists the caveman CLI as its own package, separate from the Claude Code plugin (claude plugin install caveman@caveman), so installing only that plugin or only the skill folder is not enough for this skill. Run caveman learn once afterward to see the report this skill acts on.

From the project README. Source.

What's in the skill

When it loads

Act on a Caveman learn report - review the ranked token sinks, apply cost-lowering fixes with per-edit consent, and report what those fixes returned. Use when asked to lower an agent's token cost, what caveman has saved, to trim a heavy CLAUDE.md, or to offload re-pasted context into cavemem.

Key facts

Author: Julius Brussee
Repo: juliusbrussee/caveman
Skill folder: skills/caveman-learn
skills.sh installs: 113,657
Repo stars: 116,702
License: Apache-2.0
Last repo commit: 2026-11-01
Bundled files: 5 files besides SKILL.md: CLAUDE.md, README.md, package.json, tests/index.js, tests/skill-file.test.mjs
skills.sh audits: Gen Agent Trust Hub: pass; Socket: pass; Snyk: pass

What it includes

Read the plan: Cave Score, its four components and the ranked token sinks
Runs code · caveman learn report --json
Show scale over scanned history; never a forward projection
Runs code · caveman learn simulate <sink_id>
Trim a heavy CLAUDE.md or an unused skill, with a diff and a yes
Runs caveman learn apply <sink_id> --dry-run, proposes a diff with before and after tokens per turn, and reverts if the edit does not lower them.
Runs code · REDUCIBLE fixes
Move a re-pasted block into cavemem and leave a pointer
Stores the block with caveman mem remember, writes a pointer where it was, and checks recall works before it finishes. Reverts with caveman mem forget if recall fails.
Runs code · RECURRING_CONTEXT offload
Turn a repeated tool sequence into a skill, with a holdout test
Starts caveman learn experiment and tells you the payback is unproven. Needs at least 5 sessions per arm before any verdict.
Runs code · SKILL_DISTILLATION
Fix duplicate rules, orphan memory files, truncation and broken imports
Audits CLAUDE.md, AGENTS.md, GEMINI.md, .claude/rules and Claude Code auto memory. One edit, one yes.
Write · MEMORY_HEALTH
Report what applied fixes returned, grouped by how it was measured
Runs code · caveman learn savings

From the official docs

Permissions

Medium risk

Claude edits your agent instruction and memory files and runs the caveman CLI. Read each diff before you say yes, and keep those files in git.

What it can do

Read your data
Create and change things
Delete things
Read-only mode

Watch out for

It edits your CLAUDE.md and memory files
Each edit needs your yes and is meant to be reversible, but keep the files under version control so you can undo it. github.com
The CLI sends usage stats by default
The README says it sends command and token-count stats, not prompts or code, and that caveman telemetry off or DO_NOT_TRACK=1 turns it off. github.com
Savings figures are estimates
SKILL.md says numbers are local and never verified. Treat them as a guide, and read the confounders it lists.

Security advisories

No published security advisories as of Nov 2, 2026.

Technical details

Turning on read-only	The analyzer (caveman learn) is read-only, but this skill is the writer: it edits files only after your yes.
Talks to	The caveman CLI may send usage stats; the README says caveman telemetry off or DO_NOT_TRACK=1 turns that off
Sign-in and keys	No account or key is named in SKILL.md.
Scopes	Runs caveman learn and caveman mem commands Edits CLAUDE.md, AGENTS.md and memory or rules files, after your yes per edit Writes candidate files under ~/.caveman/candidates/ and temporary files

Questions

What does caveman-learn do?

It reads the report from caveman learn, ranks where your agent's tokens go, and applies cost-lowering edits one at a time, each only after you say yes. It then re-measures and reports what the fix returned.

How do I install caveman-learn in Claude Code, Claude or Codex?

Use the install box at the top of the page and pick Claude Code, Claude or Codex. Claude Code and Codex read skills from a folder on your machine. In the Claude app you upload the skill folder as a zip.

Does it need the caveman tool installed?

Yes. The skill runs caveman learn and caveman mem commands, so the caveman CLI has to be on your machine. The README sets it up with npx @caveman-ai/cli@latest and lists Node.js 22.13 or newer. INSTALL.md lists the CLI as a separate package from the Claude Code plugin.

Will it change my files without asking?

No. The binding rules say consent per edit, with no apply-all that hides the individual diffs. Every edit is reversible, and it must report exactly what it changed.

Does it promise a dollar saving?

No. SKILL.md says never to claim a saving it has not measured and never to present a local number as verified. Spend figures describe what the scanned window cost, not what a fix would return.

Is it safe to use?

It edits files such as CLAUDE.md after your approval and runs the caveman CLI. The README says the CLI sends usage stats by default and shows how to turn that off. Read each diff before you say yes.

Information

updated Nov 2, 2026

Repository: juliusbrussee/caveman
Folder: skills/caveman-learn

juliusbrussee/caveman skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 5

caveman README · github.com
caveman INSTALL.md · github.com
caveman-learn README · github.com
caveman-learn SKILL.md · github.com
caveman-learn on skills.sh · skills.sh