Caveman

Use Caveman to shorten a coding agent’s visible replies while keeping exact errors, commands and outcomes. Compare the same task before and after enabling the style.

Compare a terse reply with its source

Check that shortening the visible response preserves the exact error and verification status.

Which technical reply should be shortened?

Provide a saved log, diff or handoff with facts that must stay.

Copy the request into a new Floatboat conversation and provide the required material.

Download Floatboat

Your request for Floatboat

Preview and copy the request. It is not sent automatically from this page.

Reference: Caveman
https://github.com/JuliusBrussee/caveman/blob/main/skills/caveman/SKILL.md

Read the original documentation and check the current environment and dependencies. Propose any new installation or tool connection before proceeding.

My task:
Use Caveman to summarize this saved test run. Lead with the failure. Keep the exact test name, error, command and uncertainty. Do not claim the test was rerun.

Expected deliverable:
A shorter reply with the same technical evidence.

About the Caveman Skill

Caveman changes how an agent writes its replies. It cuts filler and puts the answer first while retaining technical facts and explicit language requirements. The repository licenses the Skill under Apache-2.0 from version 3.0.0 onward. Engine compression serves a separate input-side workflow. Its companion instructions cover investigation before editing and bounded refactoring. Use those for changes to the code itself.

Caveman Skill Use Cases for Concise Replies

Choose a reply that has clear technical facts to preserve. Compare what a developer can still act on before using terse mode in everyday work.

Summarize a test failure

Provide one saved test result and ask for the failing test, error and next check. Compare the normal and terse replies against the log to make sure neither omits the exact failure.

Write a small change handoff

Give the agent the same diff and acceptance checks in two modes. Keep the shorter handoff if the affected files, observed results and remaining uncertainty are still easy to find.

Reduce status-update noise

Use an approved task transcript and ask the agent to surface the decision, blocker and next action. Check that the short version still names the owner or dependency when it matters.

Return to normal mode for a delicate explanation

If a risk or rollback condition needs a detailed explanation, ask the agent to stop using Caveman. Check that the longer response still tracks the same technical evidence.

How to Check What a Shorter Reply Still Contains

Give the agent an unchanged task, then compare the two outputs. Exact error strings and the decision behind a fix should survive even when the response is shorter.

Install the Skill for your host
Follow the upstream host-specific instructions. Start with the reply-style Skill if your task is to change visible output. Invoke /caveman where supported or $caveman in Codex.
Run a paired comparison
Give the same input to a normal reply and a Caveman reply. Record the response language, errors, commands and observed result.
Inspect facts before judging length
Compare paths, commands, negations and uncertainty with the source. Stop the mode when the task needs more explanation.

Explore More Agent Skills

Read another Skill guide and check its original source before using it.

Caveman FAQ

Does Caveman compress what the agent reads?

The reply-style Skill changes visible output. The Engine has separate input compression and recovery behavior.

Will it switch my reply language?

The manifest prioritizes an explicit reply-language instruction. Technical terms and errors should remain exact.

Can I tell whether the host tracks the mode?

The Skill can report the hook’s mode when available. A host without the relevant hook can only report the mode the agent followed.

Can I turn it off during a task?

The upstream Skill documents stop caveman or normal mode to stop the style.

Can I promise a fixed token saving?

Use like-for-like output measurements on your own tasks. The Skill does not guarantee a fixed reduction in context or billed tokens.

What should I compare besides length?

Check paths, numbers, errors, negations, failed checks and the next action against the same source material.