Caveman skill, with 75,000 stars, promises to cut Copilot token use by 75%

After Copilot's pricing change, developers began hitting their monthly token limits within the first working week. Everyone in work chats is looking for a way to save, and one skill keeps coming up among the suggestions: Caveman. The repository's README promises up to 75% savings with no loss of quality. We look at how it works and whether it holds up in practice.

What Caveman is

The idea is disarmingly simple: a special system prompt instruction tells the language model to talk like a caveman, with no articles, no introductory phrases, and extremely short sentences. The principle of "speak tersely" cuts the number of output tokens in every response, and total daily consumption drops significantly. The authors claim savings can reach 75%, and without any degradation in quality.

The repository has collected more than 75,000 stars on GitHub, which looks impressive for a tool with such conceptually simple mechanics. It was the skill's viral spread that set off the wave of discussion in team chats after Copilot's pricing policy change.

Why token limits became a problem

Copilot moved to new pricing, and it hit hardest the teams that use AI assistants heavily, with dozens or hundreds of requests a day. Developers are reaching the ceiling faster than expected, and far from everyone is willing to pay extra for larger packages.

The standard ways to save have been known for a long time:

• Write short, precise prompts
• Do not pass unnecessary context in every request
• Remove repeated instructions from the conversation
• Limit the size of the context window
• Do not ask the model to "think out loud" when it is not needed

All of this works, but it demands conscious effort from every team member on every request. Caveman offers a systemic solution: add the instruction to the system prompt once, and the model switches to economy mode on its own, with no additional oversight.

How the skill saves tokens technically

At its core, Caveman is a set of instructions that force the model to compress its answers as much as possible. The key mechanic is reducing output tokens, which at most providers cost more than input tokens. Every extra paragraph, introductory phrase or repetition is real money under intensive use. If the model switches to a telegraphic style on every request, the total savings over hundreds of interactions a day can indeed be substantial.

The 75% figure is theoretically achievable with aggressive compression, especially if the tasks do not require detailed explanations or long documentation.

"Can 75,000 stars on a repository be wrong?"

With that question, the author begins practical testing.

Where the skill can fail

Skepticism is warranted here. The phrase "talk like a caveman" is vague, and models interpret it differently. In practice, apparent savings easily turn into extra costs: the model cuts important explanations, and the user asks three clarifying questions instead of one.

The telegraphic style makes code with comments harder to read. For documentation tasks, excessive compression produces unreadable results.

What counts as "loss of quality" is a separate question. If a task is done correctly from a technical standpoint, but the explanation is so compressed that the developer does not understand the logic, is that savings or degradation? The skill's authors apparently mean only tokens, not the cognitive effort of decoding the answer.

The real benefit depends on the type of task. For short questions and small code fragments, the skill works well. For complex refactoring or explaining architectural decisions, it may do more harm than good.

What this means

Caveman's viral success is a symptom of a broader trend: AI tools are getting more expensive, limits are tightening, and the community is looking for systemic ways to stay within budget. A real test will show whether the "caveman" style is worth its 75,000 stars in practice, or whether they are just impressive numbers in a README.