Caveman seems genuinely useful for workflows that are mostly machine-facing or that produce fixed-form artifacts for humans. In those cases, the degraded prose is mostly confined to exception paths, debugging, or terse status output.
It also seems useful for vibe coding, where the person driving it does not really care about the code and judges only the resulting product. In that mode, compressed prose matters much less because the bar is mostly external: does the thing appear to work. That said, I do not think this mode creates much lasting economic value. Real value usually comes from systems that remain understandable, extensible, operable, and trustworthy over time.
My objection is about deep collaborative work. When I work closely with an agent, I am not just extracting conclusions from it — I am spending hours inside its language. Prolonged exposure means I absorb patterns of phrasing, compression, rhythm, and thought. If those patterns are intentionally degraded for token efficiency, they may still bleed into my own sense of expression. I do not want a large fraction of my reading environment to consist of deliberately broken, impoverished English.
Also, “it only affects what it outputs, not the code” is a much stronger claim than the evidence supports. Their evals mainly show that the style reduces output tokens. They do not seriously prove that the technical result stays the same, let alone that the code stays the same. So the burden of proof for “same code output” is nowhere near met.
And for me this becomes even more problematic because I strongly believe projects should have .md specs in the repo, and the agent should be able to edit them. That is where caveman starts to actively worry me. The style will not stay neatly confined to disposable chat replies — it will bleed into the spec itself. And if it starts editing an existing spec under compression pressure, I worry it will collapse context: flatten nuance, drop still-valid ideas, and gradually erode the reasoning that made the spec useful in the first place.