[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1luu84zfyub5x":3},{"code":4,"message":5,"data":6,"timestamp":58},10000,"success",{"full_name":7,"position":8,"description":9,"language":10,"stars":11,"forks":12,"stars_gained":13,"age_days":14,"hn_threads":15,"one_liner":32,"summary_md":33,"topic":34,"github_topics":37,"license":48,"homepage":49,"updated_at":50,"history":51},"JuliusBrussee\u002Fcaveman",6,"🪨 why use many token when few token do trick. Viral skill + proxy for coding agents that cuts 65% of tokens by talking like a caveman.","Go",114843,6622,229,203,[16,22,28],{"id":17,"title":18,"points":19,"comments":20,"posted_at":21},47647455,"Caveman: Why use many token when few token do trick",904,366,"2026-04-05",{"id":23,"title":24,"points":25,"comments":26,"posted_at":27},49646387,"Respond terse like smart caveman. All technical substance stay. Only fluff die",2,0,"2026-09-10",{"id":29,"title":30,"points":25,"comments":26,"posted_at":31},47794436,"Caveman – why use many token when few do trick","2026-04-16","给编程 Agent 省 token：缩短回答、压缩读取内容，且不改写你自己的提示词。","## 解决什么问题\n\nAgent 会话里的 token 往往不是花在回答上：冗长的解释、整段贴回的日志和 JSON、同一条命令输出被反复重读，才是主要开销。光加一句「请简洁」到不了这些地方。Caveman 是一个技能加本地代理，装进已有编程 Agent 里，回答侧和输入侧一起压缩。\n\n## 有什么不同\n\n常见的省钱做法有两类：一是纯提示词，比如只装 `\u002Fcaveman` 技能；二是通用 LLM 网关，夹在你和 API 之间，但看不到工具调用。Caveman 接进 Agent 运行时（README 称支持 30+ 个），因此能在模型读之前压缩工具输出和日志——「整场会话输入 token 少 33.2%」来自这一层，不是靠说话风格。真正有差异的有三点：基于 ASD-STE100 受控英语的结构化短答规则，代码、命令、报错原文逐字不动；Blocks 代码模式让 Agent 保留并复用写过的脚本，而不是每次重新获取输出（README 称一次测试从 397 token 降到 76）；以及按请求自动选模型和推理强度的 Auto。它还明确不改写你的提示词，理由是外部研究显示这样会适得其反——这比多数提示词裁剪工具更有分寸。\n\n## 适合谁\n\n按量计费跑长时间 Agent 会话、或经常被测试输出和日志撑爆上下文的开发者。按席位付费但上下文压力大的团队也有收益。用 GitHub Copilot 这类按请求计费的人可以跳过——README 自己承认短回答在这种情况下不省钱。需要提示词可审计、完全确定的人也别用：一个改写输入的代理就在你的信任链上。\n\n## 上手成本\n\n要求 Node.js 22.13+，安装命令是 `npx @caveman-ai\u002Fcli@latest`，另提供 shell 和 PowerShell 一行脚本覆盖更多 Agent。Apache-2.0 许可。全程本地运行，用你自己的 key 和 Claude Pro\u002FMax 登录。风险在于：语音规则本身约 1,160 token 输入，收益取决于会话长度；只有 Auto 需要免费账号；README 里的部分数字是阶段性结果，有些模块明确标注「未测量」。外部验证倒是存在（Adobe Research 论文、JetBrains A\u002FB、Elastic 的实践文章），在这类项目里并不常见，值得纳入考量。\n\n## 点评\n\n项目 203 天，约 11.48 万 star，但趋势榜显示最近两天日增仅两百出头——4 月 HN 榜首（904 分）的势头已经转成维护期。README 里专设「诚实数字」文件并引用外部评测，说明维护者在乎结论可验证，这对一个梗味很重的省 token 项目来说很难得。65% 这个头条数字其实是几块分开测的结果之和，每块都有自己的前提。建议先免费试技能，看短回答是否合你的工作流，再决定是否让代理层进入你的工具链。",{"slug":35,"name":36},"developer-tools","开发工具",[38,39,40,41,42,43,44,45,46,47],"ai","anthropic","caveman","claude","claude-code","llm","meme","prompt-engineering","skill","tokens","Apache-2.0","https:\u002F\u002Fdocs.caveman.so\u002Fcaveman","2026-10-25T00:05:12Z",[52,56],{"date":53,"stars":54,"stars_gained":55},"2026-10-24",114818,225,{"date":57,"stars":11,"stars_gained":13},"2026-10-25",1792901979]