Caveman skill with OpenCode
Hi developer 👋 in the following video we test the Caveman skill with OpenCode. The idea is to make the agent respond more briefly, reducing token output without losing technical precision. We also test compressing natural language in files, reviewing code, and generating short commit messages.
Installation
First, install the Caveman skill in the project:
Select all the skills it offers. Caveman is simply an instruction to the LLM that says “respond telegraphically, no filler”. The LLM follows that instruction regardless of the language being used.
File compression
Test file compression with:
/caveman-compress AGENTS.md
It creates a backup AGENTS.original.md and compresses natural text without touching code blocks, URLs, or important technical terms. If satisfied, delete the backup. Git also allows you to revert.
Activating caveman lite
There are 4 modes available. Activate the lite mode:
/caveman lite
It activates for the session. You can also modify AGENTS.md to auto-start caveman in lite mode.
Testing plan and build modes
In Plan mode, ask for a code review:
In src/DoorManager.ts the use of player.isLocked is confusing since locked sounds like blocked. Could we use a term like isMouseCaptured?
The output should be more concise. Switch to Build mode and ask to implement the change. Observe the response style.
Code review
Run a full review of the project:
/caveman-review review the project code
The feedback includes nits — comments of little importance that don’t affect functionality or security.
Fix issues and commit
Correct an issue from the review and commit with:
/caveman-commit
It uses Conventional Commits conventions.
Video
In the following video you can see the complete process (Spanish audio).