Guide · Agent skill
How to use the Cavecrew Subagent Delegation agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
Cavecrew is an agent skill for Claude Code and other SKILL.md-compatible agents that teaches the main agent when to delegate work to three specialised helper subagents: cavecrew-investigator for locating code, cavecrew-builder for one- or two-file edits, and cavecrew-reviewer for diff review. This guide explains what the skill changes, how to install it, how to trigger it, and where it is a poor fit. By the end you will know how to set it up and when to expect compressed output.
Cavecrew Subagent DelegationJuliusBrussee/caveman · Apache-2.0 Download the skill Quick start
- 1Download the ZIP. Get the cavecrew ZIP from this page. It contains the SKILL.md file.
- 2Unzip into a skills folder. Place the skill in ~/.claude/skills/ for all projects, or in the project's .claude/skills/ folder for one project.
- 3Confirm the helper presets exist. Make sure cavecrew-investigator, cavecrew-builder and cavecrew-reviewer are available in your agent setup, since the SKILL.md alone only describes them.
- 4Ask a location question. Ask where a symbol is defined or used. The agent should route this to cavecrew-investigator and return a path:line list.
- 5Hand over a known edit. Give the builder an exact path and line for a change touching one or two files.
- 6Review the diff. Ask for a review of your branch or diff. The reviewer returns findings with severity and a totals line.
What the cavecrew skill makes the agent do
The skill tells the main agent when to hand a task to one of three subagent presets instead of working inline or using the vanilla Explore agent. Each preset returns its result in a compressed format, so the main conversation keeps more of its context budget.
The skill's stated reasoning is that subagent results are added to the main context as-is. A prose answer of around 2k tokens costs that much context every time, while the same finding from cavecrew-investigator comes back as a compressed path:line table.
- cavecrew-investigator: finds where a symbol is defined, what calls it, or where something is used.
- cavecrew-builder: makes a surgical edit when the scope is obvious and touches 2 files or fewer.
- cavecrew-reviewer: reviews a diff, branch or file for bugs.
- Use vanilla agents (Explore, Code Reviewer) when you want explanations, suggestions or architecture commentary.
Install the cavecrew skill
Download the ZIP from this page. It contains SKILL.md. Unzip it so the skill folder sits in ~/.claude/skills/ for all your projects, or in a project's .claude/skills/ folder for one project only.
If you use the Claude apps instead, upload the skill under Settings → Capabilities → Skills.
- Personal install: unzip into ~/.claude/skills/
- Project install: unzip into the project's .claude/skills/ folder
- Claude apps: Settings → Capabilities → Skills, then upload
Check the requirements before you rely on it
The skill file alone only describes the helpers. You need Claude Code or another SKILL.md-compatible agent with subagent support, and the three presets cavecrew-investigator, cavecrew-builder and cavecrew-reviewer must be available in your setup. If they are missing, the skill has nothing to delegate to.
The skill runs no scripts and makes no network calls, according to the listing.
- Claude Code or another SKILL.md-compatible agent with subagent support
- The cavecrew-investigator, cavecrew-builder and cavecrew-reviewer presets installed
- Optional: vanilla Explore, Code Reviewer or feature-dev:code-architect for the tasks the skill routes elsewhere
Trigger the skill with natural requests
You do not need a special command. Ask in plain language and the agent uses the skill's routing table to choose a helper. Phrase lookups as questions about location or usage, and phrase edits with an exact file and line when you have them.
The skill's own rule of thumb: if you would want the subagent's output in a third of the tokens, pick cavecrew; if you would want prose, pick a vanilla agent.
- "Where is parseConfig defined and what calls it?" routes to cavecrew-investigator.
- "List every use of the retry flag in src/." routes to cavecrew-investigator.
- "Change the timeout at src/client.ts:42 to 5000." routes to cavecrew-builder when the file is already known.
- "Review the diff on this branch for bugs." routes to cavecrew-reviewer.
- "Explain the architecture of the auth module." should go to a vanilla agent, not cavecrew.
Use the three common chaining patterns
The most common flow is locate, fix, verify. The investigator returns a list of sites, the main thread picks one or two, the builder edits them, and the reviewer audits the diff. For broad searches you can run two or three investigators in one message, each with a different angle such as definitions, callers and tests, then combine the results in the main thread.
When you already know the exact path and line, skip the investigator and hand that location straight to the builder.
- Locate, fix, verify: investigator, then builder, then reviewer
- Parallel scout: 2-3 investigators with different angles in one message
- Single-shot edit: go straight to cavecrew-builder with a known path:line
Read the compressed output formats
Each helper has a fixed output shape, so you can parse it. The investigator lists findings as path:line with the symbol in backticks, or returns No match. The builder returns a line-range summary and a verification line, or a single terminal word. The reviewer lists findings with a severity and a totals line, or returns No issues.
Cavecrew output is structured and sometimes terse to the point of cryptic. If a person will read it directly, ask for a paraphrase.
- Investigator: path:line — `symbol` — short note, then totals
- Builder: path:line-range — change, then verified: re-read OK or mismatch
- Builder terminal words: too-big, needs-confirm, ambiguous, regressed
- Reviewer: path:line: severity: problem. fix, then totals with counts by severity, or No issues.
Avoid the common mistakes
The skill names several patterns that waste turns. Do not send the builder a file you have not located yet, and do not chain the investigator into the builder for a large refactor, because the builder will return too-big. Do not ask the reviewer for general feedback, since it only returns findings.
The skill also notes that subagents switch to normal English for security warnings, irreversible-action confirmations, and any output where fragment ambiguity could be misread.
- Do not use cavecrew-builder when you do not know the file yet.
- Do not use it for a 3+ file or cross-cutting refactor; use the main thread.
- Do not expect architecture advice or prose from cavecrew-reviewer.
What people use it for
- Finding every call site of a function in a large codebase before changing its signature.
- Making a one-line config change at a known file and line, then verifying it was applied.
- Running parallel investigations for definitions, callers and tests in one message.
- Reviewing a branch for bugs as a final check before merging.
Frequently asked questions
How do I install the cavecrew skill?
Download the ZIP from this page and unzip it into ~/.claude/skills/ for all projects, or into a project's .claude/skills/ folder for one project. In the Claude apps, you can instead upload it under Settings → Capabilities → Skills.
What do I need to run cavecrew?
You need Claude Code or another SKILL.md-compatible agent with subagent support. The cavecrew-investigator, cavecrew-builder and cavecrew-reviewer presets must also be available, because the SKILL.md file only describes them.
Is cavecrew free to use?
The listing states the skill is licensed under Apache-2.0. The source does not mention pricing, so check the licence terms in the project repository for details.
Does cavecrew run scripts or access the network?
According to the listing, the skill runs no scripts and makes no network calls. It consists of a single SKILL.md file.
What is the difference between cavecrew and the Explore agent?
Cavecrew does the same job as Explore but returns compressed output, so it uses less of the main conversation's context. If you want prose explanations or architecture suggestions, the skill recommends the vanilla Explore agent instead.
Why does cavecrew output look cryptic?
The helpers answer in a terse, compressed format on purpose to save context. The skill warns that this can be hard for humans to read, so paraphrase it or ask the agent to explain it when a person needs to read it directly.
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
Cavecrew suits developers using Claude Code on large codebases who run out of context during long sessions and want quick file-and-line lookups, edits and reviews. It is a poor fit if you need explanations, architecture advice, large multi-file refactors, or readable reports, and it assumes you already have the helper presets set up.
Download the skill