caveman-manage Skill for Claude: Review Caveman Experiments and Block Unsafe Actions

caveman-manage makes your assistant review a Caveman experiment, recommend one action, and refuse to run any lifecycle change itself.

Installs: 111K; on skills.sh
Stars: 116.7K; on GitHub
Updated: Nov 2026; last repo commit

caveman-manage reads the state and results of a Caveman Cloud experiment and checks the evidence: sample sizes, quality result and guardrails such as latency, errors and cost. It proposes one action (start, approve, cancel or rollback) with the reason, and marks the execution as blocked.

The SKILL.md says the agent connection is read-only on purpose, and that the skill must never run a lifecycle change even if you approve it. After you act yourself, it re-reads the experiment and reports what the server says. It needs a Caveman login and project. It bundles no scripts.

Good for

Teams that run Caveman experiments and want an evidence check first
Reviewers who approve changes to LLM routing or prompts
Anyone who wants an assistant that recommends but does not act

Install caveman-manage

Claude Code

Copy this prompt.

Install the caveman-manage skill in Claude Code. Run this command:

npx @caveman-ai/cli@latest

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Requires Node.js 22.13+.

This skill drives the caveman command line tool, so the skill file alone is not enough. The README lists it under the commands you use in your agent after the full install. Run caveman login before first use. Pick one route, not both.

From the project README. Source.

What's in the skill

When it loads

Inspect Caveman Cloud's experiment lifecycle and block unsafe execution. Use when asked to start, approve, cancel, promote or roll back a Caveman experiment.

Sections in SKILL.md

Non-negotiable gates
Step 1 — Load project and experiment
Step 2 — Evaluate evidence
Step 3 — Propose one action
Step 4 — Block unsafe execution
Step 5 — Re-read after external operator action

Key facts

Author: juliusbrussee (GitHub)
Repository: juliusbrussee/caveman, folder skills/caveman-manage
Installs on skills.sh: 110,982 (as of 2026-11-02)
Repo stars: 116,702 (whole repository, as of 2026-11-02)
License: Apache-2.0
Last repo commit: 2026-11-01
Files in the skill folder: 1 file: SKILL.md
skills.sh audits: Gen Agent Trust Hub: pass, Socket: pass, Snyk: pass

What it includes

Get the experiment and its results
Uses caveman_context and caveman_experiment_get, or the CLI experiments list, show and results commands.
Network · Step 1: Load project and experiment
Report state, samples, quality, guardrails, evidence cost
Marks results pending, failed, promotable or active, and says "evidence incomplete" when a required field is missing.
Read · Step 2: Evaluate evidence
Start, approve, cancel or rollback, with the reason
Shows the recommendation and id and marks execution as blocked until the server enforces the lifecycle and evidence gates.
Step 3: Propose one action
Never emit or run a lifecycle command
Explains why the CLI and MCP agent surfaces only read experiments.
Step 4: Block unsafe execution
Report what the server shows after your own action
Reads the detail and results again and reports the server-observed state, not your intent.
Network · Step 5: Re-read after you act

From the official docs

Permissions

Low risk

It reads experiment data from your Caveman project and is written never to run a lifecycle change. Check the project scope before you start.

What it can do

Read your data
Create and change things
Delete things
Read-only modeHow to turn it on

Watch out for

The install route changes your agent setup
The README says the full Caveman install finds your agents and logins, shows what it will change and waits for you to press Enter. The Claude Code plugin route auto-starts caveman in every interactive session. The README's Privacy section says the CLI sends usage stats by default (commands run, token counts, a random install ID, OS and IP, not prompts, code or file paths); `caveman telemetry off` or DO_NOT_TRACK=1 turns it off.
It reads your Caveman project data
It uses your Caveman login to read experiments and results. Check that the right project is selected.
It will not run lifecycle actions
The SKILL.md says to never run start, approve, cancel or rollback, even after you say yes. You run any action yourself.

Technical details

Turning on read-only	The SKILL.md says current agent MCP is read-only and the skill must never run a lifecycle mutation, even after user approval.
Talks to	Caveman Cloud, through MCP tools or the caveman CLI

Questions

What does caveman-manage do?

It reads a Caveman Cloud experiment, checks whether the evidence supports starting, approving, cancelling or rolling it back, and recommends one action. It does not run the action.

How do I install caveman-manage?

Use the install box at the top of the page. The skill needs the caveman command line tool or its MCP tools and a login, so the box lists the full Caveman setup first. Then run caveman login.

Can caveman-manage approve or roll back an experiment for me?

No. The SKILL.md says never to run a lifecycle change, even after you approve, because the server does not yet enforce every state and evidence check at once. You run the action yourself, then ask it to re-read the result.

Does it need an account?

Yes. It needs a Caveman login and a project. It stops if login, project, experiment or results are unavailable.

Is caveman-manage safe?

It is written as read-only and fails closed on unknown states and server errors. It still uses your Caveman login to read project data.

How is it different from caveman-evidence-review?

caveman-manage focuses on one experiment and whether an action is justified. caveman-evidence-review looks at cost, traces and savings across the project. Both are read-only.

Information

updated Nov 2, 2026

Repository: juliusbrussee/caveman
Folder: skills/caveman-manage

juliusbrussee/caveman skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 3

caveman README · github.com
caveman-manage SKILL.md · github.com
caveman-evidence-review SKILL.md · github.com