Skip to main content
Caveman integrates with your agent through three mechanisms: response skills that change how the model writes, lifecycle hooks that run during agent events, and native plugins for agents without hook support. Each mechanism has different permissions and failure behavior. This page covers every built-in skill, how to install and remove them, and what to check before you trust an integration.

​
Response skills

Response skills are instructions that tell the agent to communicate with less filler while preserving technical terms, code, error text, and necessary ordering. They affect generated prose. They do not compress request context or prove a token-reduction percentage. Install the main skill through a supported skill installer:
Preview remote skill content before installation when the source is unfamiliar.

​
Voice modes

Three sibling voice skills ship with Caveman. Each has a different output contract and can be switched at any time. /caveman ultra and /caveman wenyan still work as aliases. Auto-clarity: every skill drops to normal prose for security warnings, irreversible-action confirmations, step sequences a fragment could scramble, and a user who repeats a question. Resumes after the clear part. Claude Code and the standalone OpenCode plugin read stored mode state. Other hosts answer from the conversation and mark it (not tracked by this host). Want Claude Code sessions to start with normal prose? Set {"defaultMode":"manual"} in .caveman.json for one project or ~/.config/caveman/config.json for your user. You can also set CAVEMAN_DEFAULT_MODE=manual. Then /caveman activates the voice and /ultracave selects ultracave. Stop persists across compaction and resume; a fresh session starts off again. This startup policy is Claude Code-only: OpenCode keeps its caveman default because the installer also supplies static activation rules.

​
Example output

Question: “Why does my React component re-render?” Normal prose:
Your component re-renders because you create a new object reference each render. Wrapping it in useMemo will fix the issue.
caveman:
New object ref each render, so React re-renders. Wrap the prop in useMemo.
ultracave:
Inline object prop, new ref, re-render. useMemo.
megacave:
每繪新生對象參照，故重繪；以 useMemo 包之則免。

​
Focused skills

Focused skills perform a single task with a narrow contract. They do not change the global voice mode. caveman-stats mixes direct session counts with explicitly estimated baseline and rule-overhead fields. Estimated net can be negative. See evidence basis on each line instead of treating whole output as measured. caveman-compress preserves original in data-directory backup and validates headings, code blocks, URLs, paths, and list structure. Those checks do not prove semantic equivalence. Review diff before replacing high-stakes instructions. See Exploration and delegation for explorer and Cavecrew boundaries.

​
Hooks

The hook package supports agent lifecycle events including:
  • session start
  • user prompt submission
  • command or pre-tool execution
  • status-line updates
Hooks can add mode reminders, expose recovery context, and compact command output. They should not change user commands or hide an execution failure. Hook state uses atomic writes and restrictive file permissions. Symlink checks prevent a project-controlled link from redirecting trusted state writes. Install supported native hooks with setup commands documented in Agent wrapping.

​
Native plugins

Some agents do not expose equivalent hooks. Their profiles use native plugins:
  • Hermes Agent plugin
  • OpenClaw plugin
  • OpenCode plugin
OpenCode integration listens to native session and prompt events and can use project AGENTS.md instructions. Plugins remain subject to host agent’s plugin permissions and version compatibility.

​
Browser extension

The MV3 extension adds a visible Caveman Mode directive to outgoing messages on supported ChatGPT, Claude, and Gemini web interfaces. It runs locally and does not send analytics or prompt content to Caveman services. Extension injection is visible before send. It does not silently rewrite an already-sent message. Browser site markup changes can break selectors, so extension tests cover supported hosts and fail without submitting content. Build and test:
Load generated unpacked extension through browser extension developer mode.

​
Trust checklist

Before installing a skill, hook, or plugin:
  1. Read source and requested permissions.
  2. Confirm repository and commit or release source.
  3. Check files it will modify.
  4. Verify removal command or manual rollback path.
  5. Test in a disposable project.
  6. Avoid granting shell or browser access beyond feature need.
Skills are instructions and may influence model behavior. Hooks and plugins are code and can interact with local files or commands according to host permissions. Treat them accordingly.

​
Removing integrations

Use agent-native setup removal where available:
Remove skill packages through installer or agent skill manager, and remove browser extension through browser settings. Review agent config for stale endpoint variables afterward.