Guide · Agent skill
How to use the Caveman Explore agent skill (caveman-explore)
Updated · written with AI from the project's own documentation and checked against our catalogue data
caveman-explore is an agent skill that gives an AI coding agent a read-only helper for finding where code lives in a repository. It is licensed under Apache-2.0 and is aimed at developers working in large or unfamiliar codebases. After this guide you will know what the skill does, how to install it, how to trigger it with plain requests, and when to skip it.
Caveman Explore Repo LocatorJuliusBrussee/caveman · Apache-2.0 Download the skill Quick start
- 1Download the ZIP. Get the caveman-explore ZIP from our page. The licence is Apache-2.0.
- 2Choose an install location. Use ~/.claude/skills/ for all projects or .claude/skills/ inside one repository.
- 3Unzip the skill. Extract the archive so SKILL.md sits in the skill folder.
- 4Or upload it in the Claude apps. Go to Settings → Capabilities → Skills and upload the skill there.
- 5Confirm your runtime. Make sure your agent supports SKILL.md subagents with Read, Glob and Grep, and can use the haiku model setting.
- 6Ask a location question. Describe what you are looking for in plain language when you are orienting or searching broadly.
- 7Read the cited ranges. Open the returned file paths and line ranges yourself to confirm the code before acting on it.
What is an agent skill, and what does caveman-explore do?
An agent skill is a folder containing a SKILL.md file that tells an AI agent how to handle a certain kind of task. This one defines a subagent named FastContext that only searches and reads code.
When a main agent needs to orient itself in an unfamiliar project or locate code spread across many files, it hands the question to this explorer. The explorer runs several searches and reads in parallel, then replies only with file paths and line ranges. Its bulky search output stays out of the main agent's working context.
- Tools: Read, Glob and Grep only, so it cannot edit files or run commands.
- Model setting: haiku, chosen for speed and low cost.
- Output: one citation per line in the form path:START-END plus a short reason.
- It does not explain code, propose fixes or make changes.
Install caveman-explore from the ZIP or the Claude apps
Download the ZIP of the skill from our page. To install it for yourself across all projects, unzip it into your ~/.claude/skills/ folder. To limit it to one project, unzip it into that project's .claude/skills/ folder.
If you use the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.
- Personal install: ~/.claude/skills/
- Project install: .claude/skills/ inside your repository
- Claude apps: Settings → Capabilities → Skills, then upload
Check the requirements before you rely on it
The skill needs an agent runtime that supports SKILL.md subagents with the Read, Glob and Grep tools. Claude Code, or a compatible agent that can use the haiku model setting, fits this description.
The package includes package.json, SKILL.md and tests/skill-file.test.mjs. The skill does not run scripts and does not use the network.
- Runtime that supports SKILL.md subagents
- Read, Glob and Grep tools available
- Ability to use the haiku model setting
How do you trigger caveman-explore with natural requests?
The skill's description says to use it for cold-start orientation, broad cross-file localization, or when a direct search failed. Ask your agent in those situations and it can delegate the search to this explorer.
The description also says to skip it when the exact file or symbol is already named. In that case, point the agent directly at the file, which is simpler.
- Use it: "Where in this repo is the request retry logic handled?"
- Use it: "I'm new to this codebase. Where do the database migrations live?"
- Use it: a previous search for a string came back empty and you need broader hypotheses.
- Skip it: "Read src/router/pick.go and explain pick()."
What does a good result look like?
The explorer returns an evidence block and nothing else. Each line gives a file path, a line range it actually read, and a reason. If nothing relevant is found, it replies with the single line saying so, which is more useful than a guess.
The main agent then reads the cited ranges itself. Treat the citations as pointers to check, not as final answers.
path/to/file.ext:START-END reason it is relevantSafety notes: what the skill can and cannot touch
Because the tool list is limited to Read, Glob and Grep, the explorer cannot write files or execute commands. The skill does not run scripts and does not use the network, according to its listing.
Read access still means it can see whatever files your agent can read in the repository, so avoid pointing it at directories that contain secrets you would not want an agent to open.
- No script execution
- No network use
- No edit or shell tools
Tips for getting better locations
Name the feature, error message or string you remember. The skill favours parallel searches for symbols, strings and path patterns, so concrete clues help.
Give a clear scope. Say whether you want the answer for a whole repository or one area, and skip the skill once you already know the file.
- Quote exact error text or identifiers when you have them.
- Mention the language or folder if you know it.
- Use direct reads for known files; the skill's own guidance says that is simpler.
What people use it for
- Orienting in a large repository you have never worked in before.
- Finding where a setting, error string or feature flag is defined across many files.
- Locating related tests, configuration and implementation files in one pass.
- Recovering after a direct search returned nothing by getting broader search hypotheses.
- Delegating search from a main agent to a smaller, cheaper model in a multi-agent setup.
Frequently asked questions
How do I install the caveman-explore skill?
Download the ZIP from our page and unzip it into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. In the Claude apps, you can upload it under Settings → Capabilities → Skills.
What are the requirements for caveman-explore?
You need an agent runtime that supports SKILL.md subagents with the Read, Glob and Grep tools, such as Claude Code or a compatible agent that can use the haiku model setting. The skill does not run scripts and does not use the network.
Is caveman-explore free to use?
The skill is licensed under Apache-2.0, as listed on this page. Check the licence file in the package for the full terms.
Can caveman-explore edit my code or run commands?
No. Its tools are limited to Read, Glob and Grep, and its instructions say it never edits files, runs commands or proposes a solution.
When should I not use caveman-explore?
Skip it when the exact file or function name is already known, because the skill itself says to skip it then and a direct read is simpler. It is also not the right tool if you want explanations, fixes or code changes.
How does caveman-explore compare with a plain search?
A plain search runs in your main context and returns its full output there. This skill runs several searches and reads in parallel and returns only a short list of path:line citations, which keeps the main agent's context smaller.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
caveman-explore suits developers in large or unfamiliar repositories, and teams whose main agent hands search to a smaller model. If you need explanations, fixes or edits, or you already know the exact file, pick a different tool or read the file directly.
Download the skill