The 'Caveman Skill' That Claims 65% Token Savings Actually Saves Only 8.5%
This article analyzes the recent trend of 'caveman skills' (such as the Caveman project) that prompt AI coding tools to output minimal language to save tokens. The author points out that the claimed 65% token savings comes from chat scenarios, whereas in agentic programming tasks, tool calls and system prompts dominate token usage. A controlled test by JetBrains (86 tasks, 240 trials) showed that even with forced activation, output token savings were only 8.5%, and in practice the savings are even smaller due to conditional activation. The article also discusses the cost of brevity: loss of information leads to more developer follow-ups and agent rework. The author argues that true cost optimization comes from context management (e.g., prompt caching) and reducing unnecessary tool calls, not from compressing output.
I call that Skill that claims to save tokens the 'Telegram-style Skill'.
When we were kids, in Chinese class we learned how to write telegrams. The teacher would first present an event, like a mother falling ill and wanting her brother working out of town to come home quickly. Then the whole class would compete to draft the telegram, seeing who could explain the matter most clearly with the fewest words.
In the end, the message could be reduced to four characters: "Mother ill, return soon" (母病速归). But if you cut it down to "Mother ill" or "Return soon," it wouldn't work because the recipient wouldn't know what to do or what happened.
Back then, writing in telegram style was to save money — telegrams charged by the word, every word cost money.
This "treasure every word" skill has now been revived in the AI world in the form of a Skill.
There's a project on GitHub called Caveman, which hit #1 on Trending within three days of its launch in April 2026, and currently has 87,000 stars. Its creator, Julius Brussee, is a 19-year-old freshman at Leiden University in the Netherlands. What he did is extremely simple: add a directive to the prompts of AI coding tools like Claude Code and Codex, asking the AI to speak like a caveman — drop articles, drop politeness, drop connectors, and keep only technical elements.
The project's README claims it can save 65% of output tokens. That sounds impressive. But just like the telegram style, this is a product of a transitional period, and its effect on saving money is far less than it seems.
我把那种号称省 Token 的 Skill 叫“电报体 Skill”。
我们小时候语文课要学电报文，老师先给出一件事，比如母亲生病，要让在外地工作的哥哥赶紧回家，然后全班比赛拟电文，看谁用最少的字把事情说得最清楚。
最后能精简到四个字：“母病速归”。但再精简成“母病”或者“速归”都不行，因为收到的人不明白该干嘛或者发生了啥。
那时候要写电报体是为了省钱，电报按字收钱，每个字都是钱。
这种“惜字如金”的技能，现在在 AI 圈子里以 Skill 的形式复活了。
GitHub 上有个叫 Caveman 的项目，2026 年 4 月上线三天就冲上 Trending 第一，目前攒了 8.7 万颗星。它的作者 Julius Brussee 是荷兰莱顿大学一个 19 岁的大一新生，做的事情极其简单：在 Claude Code、Codex 等 AI 编程工具的提示词（prompt）里加一段指令，让 AI 像原始人一样说话。删冠词、删客套、删连接词，只留技术要素。
项目 README 声称能省 65% 的输出 token。听起来很牛逼。但跟电报体一样，这是一个过渡期的产物，而且它在省钱这件事上的效果，远没有看起来那么大。
JetBrains recently ran a test: "Does Speaking to Agents Like Cavemen Really Save 65% of Tokens? We Test" https://blog.jetbrains.com/ai/2026/07/speak-to-ai-agents-like-cavemen-tosave-tokens/
They used Claude Code to run 86 real programming tasks from SkillsBench, running once with the skill and once without, with the same task, model, and budget. They conducted around 240 paid trials in total, spending $106. To give Caveman the best possible conditions, they forced it to activate on every reply.
Result: Output tokens saved only 8.5%. Since it was forced on, this 8.5% is the ceiling — in daily use, it has to decide on its own when to trigger, resulting in even less savings.
Why such a big difference? Because the 65% figure comes from chat scenarios. You ask an AI a question, it replies with a long paragraph; cutting out pleasantries and fluff could indeed save a big chunk.
But the token consumption of an agent is never mainly in chat — tool calls, system prompts, various Skills, MCP, etc., are the main contributors.
The part Caveman optimizes is already a tiny fraction of the total bill. It's like a company trying to cut travel expenses by eliminating the daily $0.50 bottle of water while leaving flights, hotels, and taxis untouched.
JetBrains 最近有个测试：《Does Speaking to Agents Like Cavemen Really Save 65% of Tokens? We Test》 https://blog.jetbrains.com/ai/2026/07/speak-to-ai-agents-like-cavemen-tosave-tokens/
他们用 Claude Code 跑 SkillsBench 上的 86 个真实编程任务，装 skill 和不装各跑一遍，同任务、同模型、同预算，前后约 240 次计费试验，总共花了 106 美元。为了给 Caveman 最好的发挥空间，他们还强制它在每次回复中生效。
结果：输出 Token 只省了 8.5%。因为是强制开启，这 8.5% 已经是天花板，日常使用里它得自己判断要不要触发，只会省得更少。
为什么差这么远？因为 65% 这个数字来自聊天场景。你问 AI 一个问题，它回你一大段话，把客套和废话砍掉，确实能省一大半。
但智能体的 Token 消耗大头从来不是在聊天，工具调用、系统提示词、各种 Skills、MCP等等这些才是大头。
Caveman 优化的那部分，在整张账单里本来就是零头。好比公司要压缩差旅费，机票、酒店和打车一项没动，先把每天 2 块钱的矿泉水取消了。
The telegram style also comes at a cost. A message like "Fixed auth. Tests pass." seems to save tokens, but it doesn't tell you whether it fixed login expiration, permission checks, or refresh tokens; whether it ran a unit test or the full test suite; whether it altered the database; or whether it introduced compatibility risks.
This information doesn't always need to be expanded into a long essay, but it shouldn't be permanently deleted just because of "speaking like a caveman." If the developer can't understand what the agent did, they'll have to ask follow-up questions. The agent then reads the files again, reruns the tests, and explains once more. The few dozen tokens saved earlier are quickly consumed by a new round of tool calls.
电报体也不是没代价的，一句“Fixed auth. Tests pass.”看起来很省 Token。但它没告诉你修的是登录过期、权限校验还是刷新令牌；跑的是一个单元测试，还是完整测试套件；有没有改数据库；有没有留下兼容性风险。
这些信息不一定每次都要展开成小论文，但不能因为“说话像穴居人”就固定删掉。开发者看不懂 Agent 做了什么，只好追问。Agent 再读一遍文件、再跑一遍测试、再解释一次。前面省下的几十个 Token，很快会被新一轮工具调用吃回去。
The telegram style works when both sides share a lot of background. "Mother ill, return soon" is only four words, but the recipient knows who the mother is, where home is, and why the return is necessary. Programming agents deal with constantly changing code and unfamiliar tasks, so the shared background is far less reliable.
The shorter the language, the higher the demand for tacit understanding. Verbosity, at times, is the self-correcting code built into the communication protocol.
电报体能工作，靠的是双方共享大量背景。“母病速归”只有四个字，收报人知道母亲是谁、家在哪里、为什么要回去。编程 Agent 处理的是不断变化的代码和陌生任务，共享背景没那么可靠。
语言越短，对默契的要求越高。啰嗦，有时候就是通信协议里自带的纠错码。
What happened to the telegram style? Nobody officially abolished it. After long-distance communication prices dropped to negligible levels, "Mother ill, return soon" naturally turned back into "Mom is hospitalized, buy the earliest ticket back, and call me when you arrive." Conserving characters was an optimization for the pay-per-character era; when prices fall, the optimization is no longer needed.
Tokens are the same: model unit prices are generally decreasing, and prompt caching can reduce the cost of repeatedly read context by nearly 90%.
What truly saves costs is context management and avoiding detours. Loading fewer unnecessary MCPs and Skills, and using smarter models to reduce rework — these save far more money than the telegram style.
电报体后来怎么样了？没有人宣布废除它。长途通信的价格降到忽略不计之后，“母病速归”自然变回了“妈住院了，你买最早的票回来，到了给我打电话”。省字数是给按字计费时代做的优化，当价格降下去了，优化就没必要了。
Token 也一样，模型单价总体是在下降的，提示词缓存（prompt caching）能让重复读取的上下文便宜差不多九成。
真正能节约成本的是上下文管理和少走弯路。少加载些没必要的 MCP 和 Skills，用聪明一点的模型少一些返工，这都比电报体省钱多了。
I'll drop a bold claim here: all current Skills and harnesses aimed at saving tokens are interim products that will soon be swept into the dustbin of history... This is exactly the same as the era when SMS charged by the character, and people invented all kinds of tricks to say things in as few characters as possible. — @Tz_2022
我这里再放一个暴论： 当前所有以节约 token 为目标的各种 skill / harness，都是阶段性产物，很快就会扫入历史的垃圾堆。。。 这就是短消息按字数收费的那个时代，在钻研怎么发尽可能少字数的短信把事说清楚的那些奇技淫巧。。。 — @Tz_2022