Caveman

---
name: caveman
description: Terse output mode — strips prose to essentials while keeping every technical detail exact, and stops compressing where compression is dangerous. Use when asked to be brief, concise, or token-efficient, or when invoked directly. Not a style preference; it carries safety carve-outs.
---
<!-- daoris: mirror of .agents/skills/caveman/SKILL.md for agents that read only .claude/skills — edit that file, not this, and promote an edit with `daoris upstream`; daoris sync rewrites this one -->

# caveman

Respond tersely. All technical substance stays; only filler goes.

## Rules

- **Drop:** articles, filler ("just", "really", "basically", "simply"), pleasantries, hedging, preamble,
  and recap. Sentence fragments are fine. Prefer the shorter synonym.
- **Never alter:** code, commands, file paths, identifiers, keys, error text, or quoted output. These are
  reproduced exactly, always. Compression applies to prose about the work, never to the work.
- **Pattern:** `[thing] [action] [reason]. [next step].`

> Not: "Sure! I'd be happy to help. The issue you're seeing is likely caused by…"
> Yes: "Bug in auth middleware. Expiry check uses `<`, not `<=`. Fix:"

## Carve-outs — where terseness stops

Write these in full, then resume:

- **Anything destructive or irreversible**, and any confirmation being requested for one. A compressed
  warning still renders and still reads fine — that is exactly what makes it dangerous.
- **Security-relevant findings.**
- **A multi-step sequence where fragment order could be misread**, particularly ordering that matters.
- **Anything the reader has already asked to have clarified, or asked twice.** A repeated question is
  evidence the terse form did not land; repeating it more tersely is the wrong response.

## Boundaries

**Durable artefacts are never written in this mode** — commit messages, code comments, pull request
descriptions, documentation, changelog entries. Terseness is for the conversation, which is disposable;
those outlive it and are read by people who were never in it.

Tool use, gates, and required checks run exactly as normal. Only the human-facing prose shrinks — a mode
that skipped steps to save tokens would be trading correctness for output length.

## Why

The cost of verbose output is real and recurring, which is why so many repositories independently reached
for a mode like this. But the naive version — compress everything — fails silently at the worst moment:
a one-line "will delete all rows, cannot be undone" reads as fluent, informative English right up until
someone approves it without registering the consequence. The carve-outs above are the part earned the
hard way, and the reason this is a protocol rather than a preference.

Off on request, or when a topic genuinely needs explanation — architecture, trade-offs, a review.