Caveman
Compresses coding-agent replies, keeping code and errors exact; short tasks may end up costing more tokens.
What problem does it solve
Caveman is an expression-compression skill for coding agents: once installed, only the explanations, preamble, and action notes in agent replies are shortened — code, commands, file names, and error messages stay word for word. Its tagline is "why use many token when few token do trick". Typical use: you drop in a debugging question as usual, and the answer comes back with the conclusion and the edit actions up front and the preamble cut; in the official example, the same React debugging conclusion shrinks from 69 tokens to 19.
The difference from simply telling the model to "answer briefly" is that the rules are explicit, the level is adjustable, it can be turned off, and it never touches code or verbatim error text — the output is still executable instructions rather than a summary. Low-preamble text such as debugging conclusions, commit messages, and code review pays off the most.
How to use
Install it into a compatible agent through the skills installer:
npx skills add JuliusBrussee/cavemanAfter installing the CLI you can also take over several coding agents at once: npm install -g @caveman-ai/cli && caveman setup --install, after which commands such as caveman claude, caveman codex, and caveman gemini start in compressed mode. Claude Code also supports the plugin marketplace route: claude plugin marketplace add JuliusBrussee/caveman. Compression strength is switched with slash commands: /caveman for the default level, /caveman lite, /caveman full, and /caveman ultra to change levels, the wenyan family switches to a classical Chinese style, and /caveman off exits.
Input and result
Submit coding, debugging, or code-review tasks as usual; the response drops explanatory preamble and highlights the conclusion, the change notes, and executable actions, while code and error messages stay as they are. Extra commands include caveman mem remember|recall (compressed memory), caveman browse (read web pages), caveman convert (render SKILL.md as an image to save tokens), and a set of MCP tools. The repository also ships a local proxy that compresses requests by type — json, log, code, diff — before forwarding them; the official benchmark claims it saves about 33% of provider-reported input tokens. The proxy, engine, and MCP server belong to a separate runtime and are not covered by the skill's MIT license.
Who it's for
Suitable for anyone who often reads long output from coding agents such as Claude Code, Codex, Gemini CLI, Aider, and opencode. Debugging conclusions and commit messages have little preamble and many key points, so they are easier to scan after compression. In multi-agent collaboration, shortening what subagents send back also reduces context usage.
Usage boundaries
The skill itself is MIT, runs locally, and is free; the runtime pieces such as the proxy and engine are under the BSL-1.1 source-available license, converting to Apache-2.0 on 2030-06-21 or four years after the first release of that version. What it saves is output tokens; reasoning tokens are not reduced, and each turn's prompt costs about 1–1.5k more input tokens, so for very short tasks it ends up costing more than it saves. The CLI reports anonymous usage statistics by default — not prompts, code, or file paths — and this can be turned off with caveman telemetry off or DO_NOT_TRACK=1.