Caveman: Compression, Code Review, and Cost Tools for Claude Code

What is Caveman?

Caveman is a Claude Code plugin built around one idea: talk to the model in a stripped-down register and it spends far fewer output tokens while keeping the same technical content. Flowy indexes the plugin from its public repository, JuliusBrussee/caveman, which sits in the development category and carries 100,896 GitHub stars as of the last crawl. Flowy does not host or bundle the code; the real install command lives at the listing page linked at the top of this guide. The plugin ships 30 components: 7 commands, 20 skills, and 3 agents. Six of the seven commands share a name with a skill built for the same job, the command being the slash-triggered form and the skill being what Claude reaches for on its own; caveman-init is the one command with no skill twin, since it drops a one-time setup file rather than firing mid-session.

How does caveman mode compress your replies?

The caveman skill is the core of the plugin: it cuts output tokens by a measured 65 percent while keeping full technical accuracy, at one of six levels: lite, full by default, ultra, and three further wenyan levels. The caveman command sets the level directly or turns the mode off. caveman-help is a one-shot reference card for the modes and triggers, and caveman-stats reports real token usage for the session by reading the Claude Code session log directly rather than having the model estimate it. caveman-init makes the mode persistent: it drops an always-on activation rule into the repository so every IDE agent picks up the same behavior.

What else does it compress?

caveman-commit drafts a terse commit message in Conventional Commits format, a subject of 50 characters or fewer, with a body only when the "why" is not already obvious from the diff. caveman-compress applies the same idea to a file such as CLAUDE.md, a todo list, or a preferences file: it rewrites the file into caveman format while keeping the code, URLs, and structure intact, and keeps a human-readable backup of the original. caveman-review shrinks code review the same way, one line per finding covering location, problem, fix, and severity. The cavecrew-reviewer agent runs a fuller version of that review over a diff, branch, or file, tagging each line with a severity and skipping nits that don't change behavior.

What is Cavecrew, and when does it take over?

cavecrew is a skill, not an agent: it is the decision guide for handing a task to one of the plugin's three agents instead of doing it inline or running a generic explore step. cavecrew-investigator is the read-only one, returning a file-and-line table for questions like where something is defined, at around 60 percent fewer tokens than a plain exploration pass, and it refuses to propose fixes. cavecrew-builder takes the small, already-located edits: a one-file or two-file change such as a typo fix or a mechanical rename, refusing anything touching three files or more. cavecrew-reviewer, covered above, is the third agent and reviews what the builder produced.

Which skill fits which kind of engineering task?

Seven skills route a task by shape rather than compress it. caveman-explore is a read-only repository explorer for cold starts and broad cross-file localization. investigate-first diagnoses an ambiguous failure or a performance regression before any edit, producing evidence-ranked hypotheses. Once the cause is known, safe-refactor restructures code while preserving its behavior, surgical-patch fixes a bug at the narrowest layer that can carry it, and lean-build handles new feature work under a strict scope and an explicit stop condition. migration covers reversible schema, data, API, protocol, configuration, or dependency changes, and verify-and-stop is the last step, proving finished work already meets its acceptance conditions without expanding it.

What is Caveman Cloud?

Six more skills wire a project into what the plugin calls Caveman Cloud, a layer for measuring what LLM calls cost. caveman-setup connects a repository through the gateway so every request is measured for cost, tokens, and latency with no change in behavior. caveman-discover then finds every LLM workflow in that repository and labels it, so spend groups by what the code actually does, such as a support-reply flow or a nightly digest, instead of one anonymous bucket. caveman-evidence-review reads the results back, read-only: costs, a Cave Score and Cave Plan, workflows, traces, latency, errors, compression, routing, and verified savings. caveman-optimize turns a finding into an operator-chosen candidate change evaluated against a baseline. caveman-manage inspects an experiment's eval-gated lifecycle, covering starting, approving, cancelling, promoting, or rolling one back, and reads evidence before acting. caveman-learn closes the loop afterward: it reviews the ranked token sinks a report found, applies cost-lowering fixes such as trimming config or moving recurring context to cavemem with consent on each edit, and reports what those fixes returned.