/ 目錄 / 演練場 / Caveman
● 社群 JuliusBrussee ⚡ 即開即用

Caveman

作者 JuliusBrussee · JuliusBrussee/caveman

Caveman cuts Claude's output tokens ~65% by stripping filler words while preserving technical accuracy — faster, cheaper, still correct.

Caveman is a Claude Code skill (with ports to Cursor, Codex, Gemini, Windsurf, Cline, Copilot) that enforces a terse output style. Three intensity levels: Lite (professional terseness), Full (default), Ultra (maximum compression), plus a 文言文 / Classical Chinese mode. Ships with sub-skills: /caveman-commit (one-line commit messages), /caveman-review (single-line PR comments), /caveman-compress (trim memory files by ~46%). Benchmarks show ~65% token savings on diverse tasks with accuracy preserved.

為什麼要用

核心特性

即時演示

實際使用效果

caveman-skill.replay ▶ 就緒
0/0

安裝

選擇你的客戶端

~/Library/Application Support/Claude/claude_desktop_config.json  · Windows: %APPDATA%\Claude\claude_desktop_config.json
{
  "mcpServers": {
    "caveman-skill": {
      "command": "claude",
      "args": [
        "plugin",
        "marketplace",
        "add",
        "JuliusBrussee/caveman"
      ],
      "_inferred": false
    }
  }
}

開啟 Claude Desktop → Settings → Developer → Edit Config。儲存後重啟應用。

~/.cursor/mcp.json · .cursor/mcp.json
{
  "mcpServers": {
    "caveman-skill": {
      "command": "claude",
      "args": [
        "plugin",
        "marketplace",
        "add",
        "JuliusBrussee/caveman"
      ],
      "_inferred": false
    }
  }
}

Cursor 使用與 Claude Desktop 相同的 mcpServers 格式。專案級設定優先於全域。

VS Code → Cline → MCP Servers → Edit
{
  "mcpServers": {
    "caveman-skill": {
      "command": "claude",
      "args": [
        "plugin",
        "marketplace",
        "add",
        "JuliusBrussee/caveman"
      ],
      "_inferred": false
    }
  }
}

點擊 Cline 側欄中的 MCP Servers 圖示,然後選 "Edit Configuration"。

~/.codeium/windsurf/mcp_config.json
{
  "mcpServers": {
    "caveman-skill": {
      "command": "claude",
      "args": [
        "plugin",
        "marketplace",
        "add",
        "JuliusBrussee/caveman"
      ],
      "_inferred": false
    }
  }
}

格式與 Claude Desktop 相同。重啟 Windsurf 生效。

~/.continue/config.json
{
  "mcpServers": [
    {
      "name": "caveman-skill",
      "command": "claude",
      "args": [
        "plugin",
        "marketplace",
        "add",
        "JuliusBrussee/caveman"
      ]
    }
  ]
}

Continue 使用伺服器物件陣列,而非映射。

~/.config/zed/settings.json
{
  "context_servers": {
    "caveman-skill": {
      "command": {
        "path": "claude",
        "args": [
          "plugin",
          "marketplace",
          "add",
          "JuliusBrussee/caveman"
        ]
      }
    }
  }
}

加入 context_servers。Zed 儲存後熱重載。

claude mcp add caveman-skill -- claude plugin marketplace add JuliusBrussee/caveman

一行命令搞定。用 claude mcp list 驗證,claude mcp remove 移除。

使用場景

實戰用法: Caveman

Get crisp one-line commit messages for every patch

👤 Anyone tired of 8-paragraph AI-written commits ⏱ ~5 min beginner

何時使用: Claude writes a commit message, and it's a novel. You want it to say 'fix: off-by-one in pagination' and stop.

前置條件
  • Caveman plugin installed — claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman
步驟
  1. Commit with caveman-commit
    /caveman-commit — write a commit message for the staged changes.✓ 已複製
    → One-line conventional commit; no body unless genuinely needed
  2. Ship a branch of small commits
    For the 4 staged files, use /caveman-commit per file — atomic commits, one line each.✓ 已複製
    → git log shows 4 short commits

結果: Readable git log; bisect-friendly commits.

注意事項
  • Lost nuance when a message should be longer (e.g. breaking changes) — Override per-commit: 'write the full version, not caveman, this is a breaking change'
搭配使用: git · github

Review a PR with one-line actionable comments

👤 Reviewers (and reviewees who dislike AI walls of text) ⏱ ~15 min beginner

何時使用: You're doing a quick pass and want pointed nits, not lectures.

步驟
  1. Run caveman-review
    /caveman-review — go through the diff. One line per comment.✓ 已複製
    → Inline comments read like 'rename x → userId for consistency with api/users.ts'
  2. Escalate when needed
    On the auth file specifically, switch off caveman for a full-length security review.✓ 已複製
    → Long-form only where the stakes justify it

結果: Faster review cycles; reviewees don't dismiss comments as noise.

注意事項
  • One-liners miss nuance on architectural changes — Architectural changes aren't Caveman's job — use default mode for those
搭配使用: github

Shrink CLAUDE.md and memory notes by 40–50% without losing intent

👤 Anyone whose CLAUDE.md grew past 5k tokens ⏱ ~10 min intermediate

何時使用: Your project memory is eating context every session and you know it could be tighter.

步驟
  1. Compress in place
    /caveman-compress CLAUDE.md — keep rules and facts, cut the prose.✓ 已複製
    → File shrinks ~46%; no rule deleted
  2. Audit the diff
    Show the diff. Any rule that changed meaning?✓ 已複製
    → Per-change verdict; rollbacks if intent drifted
  3. Commit
    /caveman-commit — single-line commit for the compression✓ 已複製
    → 'chore: compress CLAUDE.md (-46%)'

結果: Same rules, half the tokens.

注意事項
  • Compression drops a subtle caveat — Always diff-review after compressing rule files; intent loss is the main risk
搭配使用: filesystem

組合

與其他 MCP 搭配,撬動十倍槓桿

caveman-skill + git

Terse commit workflow

Stage the change, then /caveman-commit, then git push.✓ 已複製
caveman-skill + github

Caveman PR reviews at scale

For the 5 PRs in my review queue, /caveman-review each; post only actionable comments.✓ 已複製
caveman-skill + claude-mem-skill

Shrink memory transcripts without losing decisions

/caveman-compress the claude-mem notes for this project; flag anything where meaning shifted.✓ 已複製

工具

此 MCP 暴露的能力

工具輸入參數何時呼叫成本
/caveman mode: Lite|Full|Ultra|文言文 Toggle mode for the current task free
/caveman-commit staged diff (implicit) At commit time; preserves staging small model call
/caveman-review PR diff Quick review pass small model call per file
/caveman-compress file path Shrink memory / rules files; diff-review before saving small model call

成本與限制

運行它的成本

API 配額
None
每次呼叫 Token 數
By design — negative; saves tokens
費用
Free (MIT)
提示
Use Ultra only on genuinely mechanical output (commits, tags); leave design discussions in default mode

安全

權限、密鑰、影響範圍

憑證儲存: None
資料出站: None — pure prompt shaping; no telemetry

故障排查

常見錯誤與修復

Output still verbose after enabling Caveman

Most clients run Caveman per-turn; put /caveman Full at the top of each session or add it to CLAUDE.md

驗證: First reply should feel distinctly terse
Caveman-commit returns multi-paragraph body

Usually triggers when the diff is genuinely large or breaks API. For true one-liner, add '—force-oneline'.

驗證: Re-run with --force-oneline
Classical Chinese mode outputs garbled

Rare; usually a font / terminal issue, not the skill. Pipe to a file and open in an editor that supports CJK.

替代方案

Caveman 對比其他方案

替代方案何時用它替代權衡
System prompt with 'be terse'You want it without a pluginInconsistent; Caveman's intensity levels + benchmarks give measurable results
Custom slash commandsYou only need one of the sub-behaviors (e.g. terse commits)Reinventing the wheel

更多

資源

📖 閱讀 GitHub 上的官方 README

🐙 查看未解決的 issue

🔍 瀏覽全部 400+ MCP 伺服器和 Skills