Home / All Skills / Workflow & Core Tools / caveman
English (EN)简体中文 (ZH)日本語 (JA)
Workflow & Core Tools · juliusbrussee/caveman

caveman

Ultra-compressed communication mode that cuts output tokens while keeping technical accuracy. Levels: lite, full, ultra and the wenyan variants. Use for /caveman, "caveman mode", "talk like caveman", "be brief" or "less tokens".

L1 Basic · Source registration onlyV1 Commit-PinnedDual-pass Audit
Author juliusbrusseeGitHub juliusbrussee/caveman ↗Stars 104500License MIT + BSL-1.1 分拆（仓库根 LICENSE 为 MIT 范围并注明 engine 等目录归 BSL-1.1，见 LICENSE.BSL / LICENSING.md；GitHub API spdx 显示 NOASSERTION）。技能本体（skill 与采用面）属 MIT。Commit 15581d1400
Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

Pure prompt-orchestrated 'output style compressor': The skill invokes no tools/APIs/scripts; its entire mechanism is providing the host model with a session-level linguistic style rule set—drop filler words, preserve technical substance, and compress according to tier.

skills/caveman/SKILL.md
Respond terse like smart caveman. All technical substance stay. Only fluff die.
The core principle lies in the first line: compression targets rhetorical fluff while strictly forbidding changes to technical content. All subsequent Rules/Intensity/Auto-Clarity/Boundaries serve as implementation details for this rule.

Session-level persistent state: Remains active throughout the entire session by default until the user explicitly exits, rather than acting as a one-off rewrite.

skills/caveman/SKILL.md
Default style for this whole session, every response, until user say "stop caveman" or "normal mode".
Persistent mode and 6 intensity levels (full being the default) are governed by switching commands in the same instruction block, through which the model self-maintains the style across the session.

Table-driven intensity levels: lite/full/ultra/wenyan-lite/wenyan-full/wenyan-ultra each feature distinct, explicit behavioral differences (what to drop, what to keep).

skills/caveman/SKILL.md
Default: **full**. Switch: `/caveman lite|full|ultra|wenyan-lite|wenyan-full|wenyan-ultra|off`.
The Intensity table in the README defines 'What change' tier by tier; ultra even includes cost rationales like 'no invented abbreviations/arrows (tokenizer benchmarks show zero token savings)'. The wenyan series contains Classical Chinese variants, compressing by character count.

Built-in 'clarity guardrails': Enforces ASD-STE100 (Simplified Technical English) specifications to prevent compression from degrading semantics.

skills/caveman/SKILL.md
Clarity register: mix ASD-STE100 Simplified Technical English into caveman, always.
STE constraints (one idea per sentence, short sentences ≤ 20 words, active voice, imperative mood, unambiguous terminology) act as the compression baseline; SKILL.md explicitly states 'Conflict between them → clarity win.'—readability takes precedence over compression ratio.

Auto-Clarity exception list: Automatically reverts to regular prose for safety warnings, irreversible operation confirmations, multi-step sequences, etc., avoiding compression-induced misunderstandings.

skills/caveman/SKILL.md
Drop caveman when: - Security warnings - Irreversible action confirmations - Multi-step sequences where fragment order or omitted conjunctions risk misread -
Formalizes 'when caveman is forbidden': high-risk scenarios require full prose. Examples demonstrate that destructive operations (DROP TABLE) require a full warning before resuming caveman mode.

Mandates standard prose for externally persisted text (commit messages, comments, issue/PR bodies, etc.); caveman only applies to conversational replies.

skills/caveman/SKILL.md
Persisted outside chat: write normal prose code, comments, commits, docs, issue/PR/MR/defect/ticket/bug-report text, memory files, third-party messages (/caveman-compress exempt).
Boundary design prevents the compressed style from polluting artifacts intended for other humans/systems (caveman-compress is a separate skill handling memory file compression, hence exempted).

Fidelity constraints on language and terminology: Follows the user's primary language, preserves technical terms/code/APIs/commands/error strings verbatim, and prohibits inventing abbreviations or corrupting negative semantics for token-saving.

skills/caveman/SKILL.md
Never drop not/never/no/only/except flip meaning worse than any token saved. Numbers, units exact.
Multiple rules collaborate: 'Preserve user's dominant language exactly…', 'ALWAYS keep technical terms, code, API names, CLI commands… verbatim', and mandatory retention of negatives—prioritizing semantics over token economics.

Embedded token cost rationale: Explicitly forbids arrows (→) and abbreviations because they offer zero savings under tokenizers while impairing readability—compression decisions are governed by measured token economics rather than surface aesthetics.

skills/caveman/SKILL.md
No causal arrows (→) either own token, save nothing.
The same section states 'never invent new abbreviations (cfg/impl/req/res/fn) tokenizer split them same as full word: zero token saved'; the ultra tier likewise explicitly specifies 'NO prose abbreviations… NO arrows (X → Y) measured zero token saving under tokenizer, cost decode clarity'.

2Core Capabilities

01Session-level compressed style (defaults to full, persists across turns until 'stop caveman'/'normal mode')
02Six intensity level switches: lite / full / ultra / wenyan-lite / wenyan-full / wenyan-ultra (three Classical Chinese tiers)
03Semantic-fidelity compression: Prohibits dropping negatives, keeps numbers/units exact, preserves technical terms and error strings verbatim
04Simplified Technical English (ASD-STE100) clarity guardrails and automatic clarification triggers
05Auto-Clarity: Automatically falls back to full prose in safety/irreversible/ambiguous scenarios and resumes afterward
06Multilingual fidelity: Responds in the user's primary language, compressing only the style, not the language itself
07Tool call silence discipline: Emits invocations directly without preamble or follow-up commentary (exempted only for safety/irreversible/clarification needs)
08Cross-host distribution: npx skills / Claude Code plugin / Gemini extension / native agent profile

4Recommended Governance & Evidence

Suggested Controls for Safe Operation
For Agent Runtime
Can be auto-loaded
For Administrators
Register source and version
Evidence & Triggers (File Paths & Excerpts)
  • SKILL.md: Pure prompt instructions and rule text; no script execution or external impact
Audit Notes: Pure prompt-based skill: zero scripts, zero network access, zero credential reading, zero local file writing, containing only markdown files within the directory. While compression rules encourage omissions, they are safeguarded by Auto-Clarity (falling back to full prose in safety/irreversible scenarios) and fidelity constraints on negatives, terms, and error strings. The only considerations are conversational guidance and multi-host activation scopes, with no code-level risks.

5Second-Pass Verification

  • [ok] 纯 prompt、无脚本无资产 — skills/caveman 目录 git ls-tree 仅 SKILL.md + README.md；plugins/caveman/skills/caveman 另有 agents/openai.yaml + assets/svg，均非可执行代码。
  • [ok] 会话级持久模式 + 退出词 — 原文含 'until user say "stop caveman" or "normal mode"' 与 'Default: **full**. Switch: …/off'，持久语义与边界一致。
  • [ok] 六档强度含 wenyan 变体 — README 'Six intensity levels' 表 + SKILL.md Intensity 表逐档存在，ultra 与 wenyan-* 行为差异有明确文字。
  • [ok] STE100 清晰度护栏 — SKILL.md 'Clarity register: mix ASD-STE100 Simplified Technical English into caveman, always.' 与 'Conflict between them → clarity win.' 均在。
  • [ok] Auto-Clarity 清单 — 原文 'Drop caveman when:' 下列出 Security warnings / Irreversible action confirmations / Multi-step sequences…，并有 DROP TABLE 示例。
  • [ok] 外部依赖清单为空 — 对两个文件扫描 URL/CLI/API/网络 token 零命中；技能体不含任何可点名的 CLI/API/端点。
  • [ok] 无凭证/无写入 — credential/env/write/save 类 token 零命中；唯一豁免提及 /caveman-compress 属另一技能的能力声明，不是本技能执行写入。
  • [ok] plugin 分发副本一致 — skills/caveman/SKILL.md 与 plugins/caveman/skills/caveman/SKILL.md 的 git hash-object 相同（ea8bf271…），diff 仅 README/openai.yaml/assets 的目录差异。

6Verdict & Scenarios

  • Actionable compression rules backed by cost evidence: Transforms 'be brief' into determinable drop/keep rules and rejects abbreviations and arrows based on tokenizer benchmarks.
  • Strict semantic fidelity constraints: Verbatim retention of negatives/numbers/terms/error strings + STE100 clarity priority ensure tokens are not saved at the expense of meaning.
  • Automatic exemptions for risk scenarios: Safety warnings and destructive operations automatically revert to full prose, ensuring style rules never override high-risk outputs.
  • Clean operational boundaries: Compresses within chats while using standard prose for persistent text (commits/docs/issues), preventing style leakage into deliverables.
  • Zero-dependency and easy distribution: Installs into 30+ hosts with just two markdown files in a single directory; MIT-licensed with zero engine overhead.
    • Best for: Ideal for AI coding and conversational environments sensitive to token billing and seeking concise responses (Claude Code, Codex, Gemini CLI, and 30+ agents), especially for users looking to reduce output tokens and reading fatigue during frequent, extended sessions; also suitable for engineering Q&A where code and error strings must remain uncontaminated by style.
    Not for: Not suitable for security audits, destructive operation confirmations, or tutorial explanations for beginners/external readers (Auto-Clarity steps in temporarily, but the overall style does not guarantee complete prose); not suitable for deliverables requiring exhaustive documentation (persisted text mandates regular prose, and if all responses require formal prose, it should not be enabled at all); not suitable for formal Chinese writing demanding strict Classical Chinese stylistic quality.
    Install Agent-Ready
    ① Verified Mirror 2026-09-09
    Option A · Direct DownloadDownload caveman.tar.gz
    sha256: 846b726914ba29d5…
    Option B · Copy Guide for Agent
    install.json
    Agents automatically download and verify mirrors from this JSON guide.
    ② Upstream GitHub
    GitHub Repo ↗
    Pinned to commit 15581d1400.