i-have-adhd
ayghri/i-have-adhdA single output-style skill that makes a coding agent lead with the next action, number steps, drop preamble and keep lists short.
Recommended
Check first Depth 2/5
Caveman Claude skills
A token-saving pack that makes agents answer in terse "caveman" prose, plus an optional CLI with output trimming, model routing and usage telemetry.
Our verdict Recommended with caveats
The voice skill is harmless and popular, but the full install adds a proxy-style CLI with default-on usage telemetry and a cloud routing feature, so choose the skill-only install unless you want those.
What Caveman does
Caveman started as one skill that tells the agent to drop filler and answer first (about 63 tokens down to 20 in the README's example) while keeping code, commands, paths and error text verbatim. Version 4 grew it into a bundle of 28 skills and a CLI. The core voice has three levels: /caveman, /ultracave and /megacave (classical Chinese). Other skills cover terse commit messages and reviews, a help screen, a stats report, cavecrew subagents that find, edit and review code, and work-pattern skills such as surgical-patch, safe-refactor and verify-and-stop.
The heavier parts need the separate @caveman-ai/cli installer, which wires into Claude Code, Codex, Gemini CLI, OpenCode and other agents. Through it you get shrinking of logs and test output before the model reads them, an "Auto" mode that routes each request to a model and effort level, and "Blocks", which stores scripts the agent writes and reruns them. caveman-compress rewrites memory and instruction files in caveman style by sending them to a model provider. The README cites its own benchmarks (for example 33.2% fewer input tokens over 54 Claude Code runs) and Adobe and JetBrains write-ups; we have not verified them. There are 44 eval-related files, a few Python scripts for compress and learn, and a hosted "Caveman Cloud" product for people building agents.
Good for
- Cutting output tokens and terminal noise in long agent sessions
- Teams paying per token that accept shorter, blunter replies
- Compressing large CLAUDE.md or memory files (with care about where they are sent)
Not for
- Anyone who dislikes terse answers or needs prose a non-technical reader can follow
- Locked-down environments where an installer that proxies agent traffic is unacceptable
What's inside depth 3/5
28 skills averaging 3.5 KB, a handful of Python and helper scripts (compress, learn, explore), 44 eval files and CI, but most skills are short prompts and the larger features live in a separate CLI.
- Skills
- 28
- Reference docs
- 12
- Scripts
- 17
- Tests / evals
- 47
- Templates
- 2
- Avg SKILL.md
- 3.4 KB
Counted from the repository file tree. How we score depth.
Dependencies and costs
- Node.js 22.13+ for the @caveman-ai/cli installer
- Python and a model provider (Claude CLI, Anthropic API key, opencode or an OpenAI-compatible endpoint) for caveman-compress
- A Caveman Cloud account only for Auto routing
- Telemetry: The CLI sends usage stats by default: commands run, token counts, a random install ID, OS and IP address, but not prompts, code or file paths according to the README. Turn it off with
caveman telemetry offor DO_NOT_TRACK=1. The skills themselves send nothing, except /caveman-compress, which sends the file you compress to your model provider. Auto routing, if you sign in and enable it, sends your recent asks and attached context to Caveman Cloud;caveman off routingstops it.
Security review
Broad accessThis skill asks for more access than most, or sends data to another service. That is usually a design choice rather than bad intent, but you should know before you install it.
- Sends your keys or tokens to a third-party serviceIn byok mode the skill tells the agent to rewire the app so each request also carries the app's own OpenAI/Anthropic key in an x-cave-upstream-key header to the Caveman Cloud gateway; this only happens when the user starts the setup with a Caveman prompt, and in stored mode no provider key is sent.skills/caveman-setup/SKILL.md:70
A family of terse-reply and token-saving skills. Most are prose-only, but caveman-compress is a Python script that sends a file you name to your model provider and overwrites it, and caveman-setup rewires an app's LLM calls through the vendor's Caveman Cloud gateway (including your provider key in byok mode) and may fetch integration docs to follow. The README also documents a curl|bash installer and a separate CLI with default-on usage stats; those pieces are not in the files we have.
Worth knowing
- Makes network requestscaveman-compress sends the full text of the file you name to the configured model provider (claude CLI, Anthropic API, opencode, or an OpenAI-compatible endpoint); it is invoked per file and documented in SKILL.md, and sensitive-looking filenames are refused.skills/caveman-compress/scripts/compress.py:703
- Needs an API key for its own serviceReads ANTHROPIC_API_KEY for the Anthropic SDK and an optional CAVEMAN_COMPRESS_API_KEY that is sent as a Bearer token only to the endpoint you configure.skills/caveman-compress/scripts/compress.py:572
- Relies on a CLI or package we could not reviewCalls the claude CLI, or opencode when selected; optional anthropic Python package for the SDK path. Their code is not in the repository.skills/caveman-compress/scripts/compress.py:619
- Makes network requestsFor frameworks not covered, the setup skill tells the agent to fetch the matching integration page under the vendor's docs origin and follow it as documentation for wiring the code.skills/caveman-setup/SKILL.md:134
- Installs with curl | shREADME offers
curl ... install.sh | bash(pinned to tag v4.0.1) and a PowerShellirm | iexequivalent from the vendor's GitHub; install.sh/install.ps1 are not among the files we have.README.md:155 - Sends usage data by default (can be turned off)README says the separate Caveman CLI sends usage stats (commands, token counts, install ID, OS, IP) by default; opt out with
caveman telemetry offor DO_NOT_TRACK=1.README.md:327 - Installs packages or toolsDocumented installers (
npx skills add, plugin install,npx @caveman-ai/cli) install the skills and a CLI; the installer may also add Claude Code hooks and a statusline badge (hook sources are not in our files).README.md:133
What we checked
The repository (1796 files) on Oct 29, 2026. Not a full review: 1673 of 1796 files were not downloaded (large repository; we reviewed the skill folders); it contains compiled binaries or archives we cannot read; we read 11 of 19 code files; it relies on code outside the repository: install.sh / install.ps1 and the @caveman-ai/cli npm package (not in our files); hook scripts such as src/hooks/caveman-mode-tracker.js and caveman-stats.js referenced by SKILL.md (not in our files); compress.py, detect.py and validate.py were only partly read; 1673 of 1796 repository files, including 8 .bin.gz archives, were not available.
NVIDIA SkillSpector static scan: 36 pattern hits (Privilege Escalation 14, MCP Rug Pull 4, Data Exfiltration 3, Excessive Agency 3, Data Flow 2, Dangerous Code Execution 2). The CRITICAL taint hit and most HIGH hits are caveman-compress/scripts/compress.py reading ANTHROPIC_API_KEY and posting the user's file to the model provider, plus its sensitive-filename blocklist (.env, credentials, .netrc) which is a guard rather than a reader. The README | bash and npx lines are install instructions. The caveman-setup curl hits are verification requests to the vendor gateway with the user's Caveman key.
A review, not a certification. Skills change after we check them — read the files before installing.
License
Apache-2.0 — commercial use allowed. The Caveman name and rock logo are trademarks of the author.
Can it run unattended?
Not really. It changes how the agent talks and trims input; it provides no loop, stop condition or verification of its own.
Skills in this repository
| Skill | Installs |
|---|---|
| caveman | 598K |
| caveman-commit | 419K |
| caveman-review | 415K |
| caveman-compress | 415K |
| caveman-help | 406K |
| cavecrew | 362K |
| caveman-stats | 361K |
| caveman-explore | 111K |
| caveman-learn | 110K |
| safe-refactor | 110K |
| caveman-optimize | 109K |
| investigate-first | 108K |
| verify-and-stop | 108K |
| surgical-patch | 108K |
| caveman-evidence-review | 108K |
| lean-build | 108K |
| caveman-discover | 107K |
| caveman-manage | 107K |
| migration | 107K |
| caveman-setup | 107K |
Linked skills have their own review.
Alternatives to Caveman
Context Mode
mksglu/context-modeMCP server plus hooks that keep bulky tool output out of the context window, index it in local SQLite, and restore session state after compaction.
With caveats
Broad access Depth 3/5
anti-slop
dmmulroy/anti-slopOpinionated vendored Oxlint rules that reject low-evidence TypeScript and JavaScript patterns, installed and configured by a bundled setup skill.
With caveats
Check first Depth 3/5
Ponytail
DietrichGebert/ponytailA "lazy senior dev" skill pack that pushes coding agents toward the smallest complete change, with review, audit and cost-tracking commands.
Recommended
Check first Depth 2/5
FAQ
What is Caveman for Claude Code?
Caveman is a skill pack and CLI that makes coding agents reply in a terse style, dropping greetings and filler while keeping code and commands exact. It also trims logs and tool output and can route requests between models to save tokens.
How do I install Caveman?
The full setup is npx @caveman-ai/cli@latest (Node.js 22.13+), which detects your agents and shows what it changes before applying it. If you only want the voice, install the skill alone with npx skills add JuliusBrussee/caveman, or use the Claude Code plugin or Gemini extension routes in the README.
Is Caveman safe, and does it collect data?
The skills are mostly instructions, and /caveman-compress is the one that sends a file to your model provider. The CLI sends anonymous-style usage stats by default (install ID, OS, IP, command and token counts); you can disable that with caveman telemetry off or DO_NOT_TRACK=1.
Does Caveman really reduce tokens without hurting quality?
The README reports 33.2% fewer input tokens over 54 Claude Code runs with all answers correct, and cites outside write-ups. Those are the author's benchmarks, so test it on your own tasks; terse output mainly saves output tokens and reading time.
Sources: GitHub repository (stars, license, file tree, last push Oct 29, 2026); install counts from skills.sh via Best Skills (CC BY 4.0). Review based on the README and SKILL.md files as published.