TL;DR: Caveman is a free, open-source skill that makes your AI coding agent answer in compressed "caveman-speak" — cutting about 65% of output tokens while keeping code, commands, and errors byte-for-byte exact. This Caveman guide covers the install, the best prompts, seven real use cases, and the honest math on when it actually saves money.
What Is Caveman? (And Why Everyone's Talking About It)
Caveman is a skill for AI coding agents that strips the filler out of every response. Install it once, and your agent stops writing "Sure! I'd be happy to help you with that" and starts writing "Bug in auth middleware. Token expiry check use < not <=. Fix:". The project's own tagline sums up this Caveman guide better than I can: why use many token when few token do trick.
Here's the problem it kills. Every coding agent is trained to be thorough and polite. You ask why a React component re-renders and you get four sentences of preamble before the actual answer. You paid for those tokens. You skimmed past them.
Before Caveman, the workflow was: ask, scroll past the throat-clearing, find the one line that mattered. After Caveman, the answer is that line. In the maintainer's own example, a 69-token response becomes 19 — same fix, nothing technical lost.
It works with Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, and 30+ other agents. MIT licensed, no telemetry, no backend. It climbed past six figures in GitHub stars because the premise is legible in four seconds to anyone who has watched an agent write three paragraphs to say "add a null check."
Who Is Caveman For?
This is a developer tool first, but the audience is wider than that. If you talk to an AI coding agent more than a few times a day, you are the target.
Specifically:
- Solo developers and indie hackers paying their own API bills, where output tokens are a line item they actually feel.
- Engineering teams on Claude Code or Cursor where nobody owns the agent bill and nobody has measured it.
- AI consultants and freelancers who can turn token reduction into a billable audit.
- Anyone hitting context limits mid-session — compression means more real work fits before the window fills.
- Non-English speakers. Caveman compresses style, not language. Write Portuguese, it grunts Portuguese.
If you mostly generate code rather than ask for explanations, read the honest-numbers section below before you install. This Caveman guide is not going to tell you it works for everyone.
Key Features of Caveman
Six compression levels
Switch anytime with /caveman <level>. lite trims gently, full is the default, ultra goes maximally terse, and wenyan renders answers in classical Chinese — the most meaning per token of any option. The level sticks for the session.
Code and errors stay exact
This is the design constraint that makes it usable. The skill compresses prose only. Code blocks, shell commands, file paths, URLs, and error messages pass through byte-for-byte. You are shrinking the commentary, not the payload.
Memory file compression
/caveman-compress <file> rewrites a memory file like CLAUDE.md into caveman-speak. This is the underrated feature. Real receipts from the repo: a 706-token preferences file drops to 285, a 1,145-token project-notes file drops to 535, averaging about 46% saved. Unlike output savings, this one compounds — every future session loads a smaller context.
Commit and review commands
/caveman-commit writes Conventional Commit messages with a 50-character subject cap, framed as why over what. /caveman-review produces one-line PR comments in the format L42: 🔴 bug: user null. Add guard.
Built-in measurement
/caveman-stats reports real token usage, lifetime savings, and a USD figure, writing a running total into your Claude Code statusline. Most token-saving advice asks you to take it on faith; this one ships a counter.
How to Get Started with Caveman in 5 Minutes
This is the section most people search for when they look up how to use Caveman. It's genuinely five minutes.
-
Check Node. You need Node 18 or higher. Run
node --version. That's the only dependency. -
Run the one-line installer. On macOS, Linux, WSL, or Git Bash, curl the repo's
install.shand pipe it to bash. On Windows PowerShell 5.1+, use theirm install.ps1 | iexform. It takes ~30 seconds, auto-detects every agent on your machine, skips the ones you don't have, and is safe to re-run. -
Or install for one agent. Claude Code users add the plugin marketplace and install directly. Gemini CLI installs it as an extension. Cursor, Windsurf, Cline, and Codex go through the skills registry with
npx skills add -a <agent>. The repo'sINSTALL.mdhas the full matrix. -
Turn it on. Type
/cavemanor just say "talk like caveman." On Claude Code, Codex, and Gemini it's already active from message one via a session hook — no command needed. Say "normal mode" to turn it off. -
Pick a level. Start with
full. Tryultraon a debugging session and see if you miss anything. You won't, mostly. -
Compress your memory file. Run
/caveman-compresson yourCLAUDE.md. This is the step people skip, and it's the one with permanent returns. -
Measure it. Run
/caveman-statsafter a real working session. Compare it against what you expected.
Beginner tip: if the install breaks, open your agent inside the cloned repo and say "Read CLAUDE.md and INSTALL.md, install caveman for me." The agent reads the repo and fixes its own brain.
7 Best Use Cases for Caveman
1. The daily coding loop
Install it, set full, forget it exists. Explanatory tasks save the most — the published benchmark shows 87% savings on "explain this React re-render bug" and 84% on "set up a PostgreSQL connection pool." Those are the prompts where agents are most verbose, so that's where the money is.
2. Long debugging sessions
Compression means your context window holds more actual work before it fills. A four-hour session that used to hit the wall and force a restart now runs longer on the same budget — invisible per-message, obvious by hour three.
3. Permanent memory compression
Run /caveman-compress on every file your agent loads each session — CLAUDE.md, AGENTS.md, .cursorrules, project notes. A 46% average reduction on a file loaded 40 times a month is a much better trade than a 65% reduction on one reply.
4. Subagent fleets
The cavecrew-* subagents (investigator, builder, reviewer) run roughly 60% fewer tokens than vanilla equivalents. If you orchestrate multiple agents, this is where savings compound hardest — you're multiplying the reduction across every agent in the fleet, and the main context survives longer.
5. MCP server slimming
caveman-shrink is middleware that wraps any MCP server and compresses its tool descriptions. If you have ten MCP servers loaded, their descriptions are eating context before you type a single word. Most people have never counted this cost.
6. Code review and commit hygiene
/caveman-review turns paragraph-long PR comments into scannable one-liners. /caveman-commit enforces a 50-character subject. Reviewers stop hunting for the actual issue inside the politeness. Your git log gets readable as a side effect.
7. Accuracy on hard problems
A March 2026 paper, Brevity Constraints Reverse Performance Hierarchies in Language Models, tested 31 models and found that constraining large models to brief answers improved accuracy by roughly 26 points on some benchmarks. Compression isn't only a cost play — sometimes less word really is more correct.
5 Copy-Paste Prompts for Caveman
These are the best Caveman prompts to run in your first week — they do the configuration work for you instead of leaving you to tune it by hand.
Prompt 1: Baseline Audit
Before I install anything, audit my current AI coding workflow. Ask me which agent
I use, roughly how many turns per day, and what percentage of my prompts are
explanatory (explain X, why is Y happening) versus generative (write this function).
Then estimate my monthly output-token spend and tell me honestly whether an
output-only compression layer would help me or be net-negative. Do not pitch me.
Give me the number.
Prompt 2: Install and Verify
Read the caveman README and INSTALL.md. Detect which AI coding agents are installed
on this machine. Install caveman for each using the correct per-agent method. After
installing, verify each one by checking that the skill file landed in the right path.
Report a table: agent, install method, path, verified yes/no.
Prompt 3: Memory File Compression Pass
Find every memory and context file this project loads each session (CLAUDE.md,
AGENTS.md, .cursorrules, project notes, any rules directory). For each: report
current token count, rewrite it in caveman-speak preserving all code, URLs, file
paths and commands byte-for-byte, then report the new token count and percent saved.
Show me a before/after diff for each file before you write anything.
Prompt 4: Honest Savings Calculator
Run /caveman-stats. Then account for the skill's own overhead of roughly 1-1.5k
input tokens per turn. Calculate my actual net savings: gross output tokens saved,
minus skill overhead input tokens, times my model's per-token pricing for each
direction. Give me the real monthly dollar figure, not the gross one. If it is
negative, say so plainly and tell me which workloads to turn it off for.
Prompt 5: Compression Regression Test
Build a reproducible benchmark for my workload, not the repo's. Pick ten prompts
that represent what I actually do. Run each in normal mode and caveman mode. Record
output tokens, wall-clock time, and whether the technical content was identical.
Output a markdown table and a verdict: which task types win, which lose, and where
the break-even sits for me.
Caveman vs. Manual System Prompts: Which Should You Use?
The obvious alternative is doing it yourself — adding "be concise, no preamble" to your system prompt or CLAUDE.md. That genuinely works, it's free, and it costs you zero install.
Where Caveman wins is consistency and coverage. A hand-written brevity instruction drifts as the conversation gets long, and you maintain it separately in every agent. Caveman ships a tuned prompt, a session hook, explicit rules protecting code and error text, six dialable levels, a measurement command, and identical behavior across 30+ agents from one install.
Use a manual instruction if you're on one agent and want zero dependencies. Use Caveman if you're on several, want the memory and subagent tooling, or want a number you can show someone.
How to Make Money with Caveman
1. Sell token-cost audits
Every team running coding agents at scale has a bill nobody owns and nobody has measured. Package a fixed-scope audit: measure current spend, install compression across their fleet, compress every memory file, slim their MCP descriptions, hand back a one-page before/after report. Price it against annual savings, not your hours — a team burning $2k/month will pay a few hundred for a credible plan to cut it. The honest framing, where you tell them when it doesn't work, is what closes it against everyone else selling hype.
2. Sell the compressed-config layer
The install is free; the configuration isn't obvious. There's a real product in pre-compressed agent configs — caveman-speak CLAUDE.md templates for common stacks, compressed subagent definitions, slimmed MCP wrappers for the most-used servers, and a level-tuning cheat sheet by task type. The moat isn't the compression. It's having already done it for the stack your buyer is on.
3. Content-to-offer on the virality
Caveman went viral because the premise is funny and the benefit is instant — a rare combination and a content gift. Film your real /caveman-stats number. Show the 69-vs-19-token comparison side by side. Then post the contrarian take: "it doesn't always save money, here's when" outperforms hype and builds the trust that converts. The tool is free, so you never compete with it — you sell judgment, which nobody installs with a curl command.
Frequently Asked Questions About Caveman
Is Caveman free?
Yes. It's MIT licensed and free forever, with no accounts, no backend, and no paid tier. The maintainer funds it through GitHub Sponsors. A separate product, Caveman 2, is in development to make savings verifiable across a team, with a waitlist at caveman.so.
Is Caveman safe to use?
It doesn't phone home. No telemetry, no analytics, no accounts, and zero network calls after install — the skill is a prompt and the hooks are local scripts. The only network activity is at install time, fetching from GitHub and your agents' own registries, which is documented in the repo's SECURITY.md. As with any piped-to-bash installer, read the script first if that's your policy.
What is Caveman best for?
Explanatory work. The benchmark range is 22% to 87% savings, and the top end is all "explain this" and "why is this happening" prompts. It's weakest on already-terse generative tasks like refactoring a callback, which saved only 22%. The memory-compression feature is arguably the best part and gets the least attention.
Does Caveman actually save money?
Partly, and the maintainer publishes the caveat himself. Caveman shrinks output tokens only — input and reasoning tokens are untouched, and the skill adds roughly 1–1.5k input tokens per turn. Whole-session savings therefore run smaller than the headline 65%, and on already-terse workloads they can go net-negative. The reliable wins are readability and speed; cost savings are the bonus. Run Prompt 4 above to get your real number.
Can beginners use Caveman?
Yes — it's one command and about 30 seconds. The one thing to know is that compressed answers assume more context, so if you're using your agent to learn rather than to ship, run lite or turn it off with "normal mode" for those sessions. Nothing is permanent and you can switch levels mid-conversation.
Final Verdict
Caveman is worth installing, with one condition: know what it does. It cuts about 65% of output tokens, keeps code and errors untouched, takes thirty seconds, and works across every agent you'd plausibly use. The virality is deserved.
The condition is that you read the honest-numbers section before you start posting savings screenshots. Output-only compression plus 1–1.5k tokens of skill overhead per turn means your net is smaller than the headline, and sometimes negative. Anyone telling you it cuts your bill by 65% flat hasn't measured it.
Install it if you use a coding agent daily, especially for explanatory work. Run /caveman-compress on your memory files on day one — that's the compounding win most people miss entirely. Then measure your own workload instead of trusting anyone's benchmark, this Caveman guide included.
Want the complete Caveman prompt pack + monetization playbook? I put together a full guide with 10 copy-paste prompts, all 10 use cases mapped out, the honest-math breakdown, and a step-by-step monetization playbook for selling token-cost audits. Grab it on Gumroad for $19 →
Published: 2026-09-19 | Updated: 2026-09-19
Top comments (0)