Caveman-speak prompting cuts Claude chat output 65 percent, but only 8.5 percent on agent tasks

Making Claude talk like a caveman really does cut its output, but the headline number is not the one that matters.

The trick and its own numbers

The idea is simple: ban preambles, pleasantries, articles and narration, and have the model return only the action and the result. “Me talk short. No explain. Tool first. Result first.” The technique is now packaged as a Claude Code skill with three intensity levels from lite to ultra.

The project’s own reproducible benchmarks show an average 65% reduction in output tokens across 10 chat-style prompts, with individual cases ranging from 22% to 87%. Only output tokens were counted; input tokens were unchanged.

The independent check

JetBrains ran the skill against SkillsBench in July 2026, on 86 real agentic coding tasks with Claude Sonnet 5, auto-graded by each task’s own test suite. The saving there was 8.5%.

That gap is the whole lesson. Chat replies are mostly the model talking, so compressing the talking saves a lot. Agentic coding runs are mostly tool calls, file contents and re-sent context, so trimming the prose barely moves the total. The skill also reports about 46% savings when compressing memory files, which persist across sessions, and that is arguably the more durable win, since a smaller memory file pays back on every future session that loads it.

Where the savings actually live

None of this makes the trick bad. It makes it specific. Teams whose bill comes from long chat answers can treat caveman mode as a real cut. Teams whose bill comes from agents reading a repository will find it a rounding error, because the prose was never the expensive part.

The practical step before adopting any token optimization is the same one this case illustrates: check where the spend actually goes, chat output or agent context, and match the fix to the number that dominates the invoice.