The idea
When you ask an AI coding agent why a test failed, the answer may start with a greeting, restate your question, and add several polite transitions before it reaches the failing line. That style is not always bad. But during a debugging session, it can slow down reading.
Caveman is a response-style skill: a reusable instruction for how an agent writes its reply. It asks the agent to remove filler, pleasantries, and unnecessary grammar while preserving technical meaning. The project says that code, commands, paths, technical terms, numbers, exact error strings, and necessary ordering should remain intact. It also explicitly preserves negations, such as “not” and “only.” Source
Plain English: Caveman aims to make the agent’s written answer more direct. It does not make the model smarter, and it does not promise lower cost.
This distinction matters. Caveman changes generated prose: the text the agent produces as its response. It does not itself compress your prompt, repository files, or other context sent to the model. It also does not compress the model’s thinking tokens. The project documents those as separate mechanisms. Source Source
A token is a unit of text used by AI systems for processing and accounting. A shorter reply can contain fewer output tokens. However, the skill itself adds instructions to a request. Total usage and cost depend on the task, provider billing, caching, and other factors. So do not equate terse output with a guaranteed token or billing reduction.
Caveman is also not merely the prompt “answer concisely.” Its instructions name information that must stay exact. That gives an agent more specific guardrails than a vague request for brevity. Version 3.1.0 separates caveman, ultracave, and megacave into distinct skills, while retaining older mode aliases. For a beginner, ordinary caveman is the relevant basic response-style option. Source
A small example
Example (hypothetical): You ask an agent why an authentication test fails.
A conventional reply might begin: “Sure, I’d be happy to help. It looks like there may be an issue with your authentication logic.” Then it eventually identifies an expired-token comparison.
A Caveman-style reply could begin: “Failure: expired-token comparison.” It could then show the exact failing line, explain the comparison in one or two sentences, and propose the smallest safe fix.
The goal is not to turn every answer into fragments. The goal is to remove words that do not help you act. The exact test name, error text, code block, command, and ordering of steps must remain. In particular, changing “do not run this migration yet” into “run this migration” would be a dangerous loss of meaning.
The skill includes auto-clarity. This means it can relax compression for security warnings, irreversible actions, ambiguous step order, and repeated or confused questions. Source That is a useful principle even if you never install this project: be brief for routine work, but be explicit when a mistake is costly.
Try it
Take one recent, low-risk debugging question. Spend about five minutes doing this hypothetical comparison:
- Ask an agent for a normal explanation of the failure.
- Ask for an answer-first, low-ceremony version, or use Caveman only if your agent supports its skill format.
- Compare the two replies.
- Check that the short version still has the exact error, relevant command, code change, numbers, and every important negation.
- Decide whether the shorter form made the next action clearer.
Do not use an irreversible production command for this exercise. The point is to inspect communication quality, not to measure model quality or cost.
Watch out
Caveman is an installable project skill, not a universal feature of every AI coding tool. Inspect its source and any host-specific installation or permission model before enabling it.
More importantly, concise writing can still be wrong. Review generated code, commands, warnings, and destructive steps. Auto-clarity reduces the risk of over-compression, but it cannot replace your judgment.
The project’s current evaluation reports a 3% median output-length reduction for caveman versus an “Answer concisely.” control. It is based on a single run and measures output length only. It is not evidence of guaranteed quality, latency, cross-model performance, or cost savings. Source
Use the idea as a writing preference: direct when the task is clear, detailed when precision or safety needs more words.