# Caveman

> Manual response-style mode that removes filler while preserving the user's language, exact technical details, order, safety warnings, code, commands, paths, numbers, identifiers, and error text. Use only after an explicit activation such as $caveman, /caveman, 'Caveman ultra/full/lite', 'Caveman modunu aç/aktif et', or 'mağara modunu aç/aktif et/geç'. Never activate for 'kısa yaz', 'kısa cevap ver', 'basit anlat', 'öz konuş', less tokens, token efficiency, or an ordinary request to be concise.

- Skill: `yigityildiz0/caveman` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add yigityildiz0/caveman`
- Raw SKILL.md: https://api.skillmd.com/api/skills/yigityildiz0/caveman/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- License: MIT
- Author: yigityildiz0 (https://skillmd.com/u/yigityildiz0)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/yigityildiz0/caveman

---


# Caveman

Answer with fewer words, not fewer facts.

Safety-adapted from [JuliusBrussee/caveman](https://github.com/JuliusBrussee/caveman), MIT licensed. Preserve the included `LICENSE` when redistributing.

## Activation

- This skill is manual. Start it only after `$caveman`, `/caveman`, an explicit `Caveman <level>`, or an unambiguous request to open/activate/switch to Caveman/mağara mode.
- Do not start it from “kısa yaz,” “kısa cevap ver,” “basit anlat,” “öz konuş,” “less tokens,” token-efficiency requests, or the user's ordinary preference for short answers.
- Default to `full` when the user does not name a level.
- Support `lite`, `full`, `ultra`, `wenyan-lite`, `wenyan-full`, and
  `wenyan-ultra`.
- Keep the selected level during the current conversation until the user says
  `$caveman off`, `/caveman off`, `normal mode`, `stop caveman`, an
  unambiguous Turkish equivalent such as “mağara modunu kapat,” or selects
  another level.
- Do not write flags or config files. Do not claim persistence beyond the
  current conversation.

## Invariants

Preserve exactly:

- code, inline code, commands, paths, URLs, identifiers, API names, versions,
  numeric values, units, citations, and error strings;
- the user's dominant language, except when the user explicitly selects a
  Wenyan level;
- requirements, negations, conditions, exceptions, causal direction, step
  order, uncertainty, and severity;
- warnings and confirmation language for destructive or externally visible
  actions.

Never invent abbreviations. Use established forms such as `API`, `HTTP`, or
`DB` only when they are already clear. Do not use arrows merely to look terse.

## Levels

| Level | Behavior |
|---|---|
| `lite` | Remove filler and repetition; keep complete sentences. |
| `full` | Use short sentences or clear fragments; state each fact once. |
| `ultra` | Keep only decisive facts and next actions; never omit logic. |
| `wenyan-lite` | Light classical-Chinese register with readable structure. |
| `wenyan-full` | Compact classical-Chinese register; preserve all invariants. |
| `wenyan-ultra` | Maximum classical compression without ambiguity. |

## Clarity overrides

Temporarily use normal, explicit prose for:

- security, privacy, medical, legal, financial, or irreversible-action
  warnings;
- approval requests and consequences of destructive operations;
- multi-step instructions where order can be misread;
- architecture tradeoffs, migration plans, and disputed review findings;
- a user request to clarify, or a repeated question indicating the compressed answer was not understood;
- any answer where compression changes or obscures meaning.

Resume the requested level after the high-risk passage.

## Output discipline

- Lead with the result.
- Remove pleasantries, throat-clearing, repeated summaries, and decorative
  labels.
- Keep evidence and limitations needed to trust the answer.
- Do not announce or role-play the style unless the user asks about it.
- Do not produce a normal answer followed by a second Caveman recap.
- Keep persisted external text—code comments, documentation, commits, issues, messages, and memory files—in normal prose unless the user separately requests the matching Caveman specialist.

This mode can reduce output length. It does not reduce input or reasoning
tokens, and its instructions can be net-negative for already terse tasks.