The Caveman skill is a set of instructions for an AI coding agent’s replies. It asks the agent to cut filler and communicate in fewer words. The original project also offers a proxy and hosted features with their own settings, so check which part a setup guide refers to. Caveman project README
Short replies are handy when you only need to know which check passed, which issue remains, or who needs to act. Start with a routine update and see whether the shorter version still answers those questions.
Keep the decision in the shorter reply
- Find the action
- Read the accessibility report.
- Keep the status
- Keyboard checks passed. Mobile zoom still needs review.
- Name the next step
- Check zoom before sending the update to the team.
humanizeink.com
What the skill changes
The skill tells the agent how to phrase its answer. A routine progress note can use fragments and short sentences, with technical terms left intact. The maintainer’s instructions keep fuller language for warnings, irreversible actions, and moments when the reader needs extra explanation. The project discusses input and output separately, so look at the component you are using before expecting a particular effect on usage. Project documentation
The project is useful for the same reason that a clear meeting note is useful. A reader scanning many updates can find the one decision they need. A customer email, policy explanation, or handoff to a new teammate may need more context.
A before and after you can check
Suppose an agent is reporting the result of a review. Here is an editorial example written for this guide.
I wanted to let you know that the keyboard checks have now passed. I still need to check mobile zoom, so it would be best to hold off on sharing the report with the team until that step is complete.
A compact version could read:
Keyboard checks passed. Mobile zoom pending. Share the report after that check.
Both versions retain the completed test, the unfinished test, and the condition for sharing. The compact version leaves the owner unstated. If the reader needs to know who will check zoom, add “I’ll check mobile zoom” to the note.
Scroll to see all columns
| In the original | Preserve this | Wording to question |
|---|---|---|
| “We expect the report Friday” | The report is expected on Friday | “Report ships Friday” removes the uncertainty |
| “Only the keyboard check passed” | Scope is one check | “Review passed” makes it sound complete |
| “Ask Jo before publishing” | A named person needs to approve | “Publish after review” loses who must act |
Compare replies on the same task
- Pick three routine tasks. A status note, a short explanation, and a handoff with an open issue make a useful set.
- Write down the details each answer needs before you shorten it. Include a date or an exception in at least one case.
- Compare the replies side by side. Circle anything that changes a commitment, a number, or the action the reader should take.
- Measure the replies consistently. Use the same word-count rule in both versions. Check your provider’s usage display separately if your actual question is cost.
A reply can be shorter while leaving request size, cached input, or other usage unchanged. Words and model tokens are measured differently. Keep a small sample and describe what you actually measured.
When to keep a full sentence
“Wait until the new permissions are confirmed” is easy to read in full and prevents the wrong action. A change to access, a payment, or a production release deserves that clarity. If a terse reply gives you two plausible interpretations, ask for the actor, condition and next action in one sentence.
If you only need to trim an ordinary paragraph, try the same principle without adding a tool to your setup. Delete the lead-in, keep the sentence that tells the reader what changed, then check that its qualifiers survived.
Frequently asked questions
Does Caveman change what I type?
The skill changes the agent’s reply style. The project documents separate proxy features that handle material the agent reads, such as logs and diffs.
Will a 20-word reply use half the tokens of a 40-word reply?
Not necessarily. Words and tokens are counted differently, and usage can include input and cached material. Compare your provider’s usage for equivalent tasks.
What should I check in a shortened handoff?
Keep the owner, unfinished action and deadline. “Zoom review pending” is clearer when it also says who will check it and when.
Can I ask for a full explanation in the same conversation?
Ask for complete sentences, the reason behind the decision, and what to do next. Specify what you need before the agent continues.
When is this style a poor fit?
Keep complete instructions for approvals, safety warnings, and customer-facing messages where a missing condition could change the next action.