# Caveman-Speak Won't Cut Your Claude Bill 65% > Source: > Published: 2026-08-20 13:08:11+00:00 [AI](https://sourcefeed.dev/c/ai)Article # Caveman-Speak Won't Cut Your Claude Bill 65% JetBrains measured 8.5% on real agent workloads — the savings worth chasing are on the input side. [Priya Nair](https://sourcefeed.dev/u/priya_nair) "Why use many token when few token do trick." That tagline carried [Caveman](https://github.com/JuliusBrussee/caveman) — a [Claude Code](https://claude.com/claude-code) skill that makes the agent answer in terse, article-free caveman-speak — to nearly 100,000 GitHub stars. The headline claim: a 65% cut in output tokens with technical accuracy intact. It's a great meme. It's also, for the workload most people will apply it to, mostly wrong — and the project's own documentation quietly agrees. The genuinely useful engineering in this repo is somewhere else entirely. ## The 65% number doesn't survive an agent benchmark The most rigorous outside test so far comes from [JetBrains' AI team](https://blog.jetbrains.com/ai/2026/07/speak-to-ai-agents-like-cavemen-tosave-tokens/), who ran paired A/B trials of Claude Code across the 86 tasks in SkillsBench — roughly 240 sandboxed runs, about $106 in API spend. With the skill force-activated (a best-case ceiling, since real installs rely on the agent choosing to invoke it), output tokens dropped 8.5%. Not 65. Eight and a half. The quality news is better: 8 tasks scored higher with Caveman, 10 lower, 64 tied — statistical noise. Talking like a caveman doesn't make Claude dumber. It just doesn't make it much cheaper either, because agentic sessions aren't prose. They're diffs, tool calls, file contents, and code blocks, all of which Caveman deliberately preserves byte-for-byte. The 65% figure comes from chat-style Q&A measured against an unprompted assistant — the most verbose possible baseline. Independent reproductions have found that a six-line "be brief" instruction captures most of the same savings, and the skill itself adds 1,000–1,500 input tokens of instructions per turn, which can push short sessions net-negative. To the maintainer's credit, the repo now says much of this out loud: savings are labeled "inferred" or "benchmark_counterfactual," never "verified"; the docs warn that terse workloads can lose money; and the project runs terse-control arms to separate Caveman's contribution from generic concision. That level of epistemic hygiene is rare in viral AI tooling. The 65% stayed in the headline anyway, because headlines are what get you 99k stars. ## Output tokens were never where your money goes Here's the structural problem no output-compression trick can fix: in a typical Claude Code session, output is maybe 10–30% of the bill. The dominant cost is input — the system prompt, CLAUDE.md, MCP tool schemas, and an ever-growing conversation history that gets re-sent on every single turn. A 65% cut to output prose, even if you achieved it, shaves single-digit percentages off total spend. JetBrains measured cost savings around 10% and noted they were "routinely erased by single-trial variance." If cost is your actual goal, the order of operations is boring and well known. First, [prompt caching](https://docs.claude.com/en/docs/build-with-claude/prompt-caching) — cached input tokens cost a tenth of fresh ones, and Claude Code applies it automatically, which is why "compress the system prompt" schemes often chase savings the cache already delivered. Second, trim what you re-send: a bloated CLAUDE.md and a dozen MCP servers with paragraph-long tool descriptions cost you on every turn of every session. Third, shorter sessions — `/clear` early, `/compact` deliberately. Caveman-speak sits somewhere around fifth. ## The interesting part is the proxy, not the grunting Which makes it notable that Caveman's own roadmap has pivoted to the input side. The 2.x release wraps a local proxy (`npm install -g @caveman-ai/cli && caveman setup --install` ) that intercepts provider traffic and applies per-content-type compression before tokens ever reach the API: JSON gets structurally squeezed while keeping keys and error trees, logs drop INFO noise but keep stack traces, code collapses to imports and signatures. Originals live in a content-addressed store on disk, and the agent can pull back full bytes via a retrieval tool if the compressed view isn't enough — a sensible lossy-with-recovery design that echoes what context-management layers inside agent frameworks are converging on anyway. The pinned benchmark claims 33.2% fewer provider-reported input tokens, which, unlike the 65%, targets the expensive side of the ledger. There's also a genuinely clever, slightly unhinged "pixel mode" that renders dense text to PNG and feeds it to the vision path, arbitraging text-token pricing against image-token pricing. And the companion tools — one reviewer measured the CLAUDE.md compressor cutting config files by roughly 46%, a one-time cost paid back every session — attack exactly the recurring input costs that matter. Caveats stack up fast, though. The proxy core is BSL-licensed, not MIT — only the skill and CLI shells are. Every compression layer between your agent and the API is a new failure mode: a mangled JSON payload or an elided function body that mattered turns a cheap session into an expensive debugging session. And the proxy's numbers are still the vendor's own; nobody has done a JetBrains-grade audit of the input side yet. I'd treat 33.2% as a claim, not a fact. ## Verdict As a cost story, Caveman is hype — a chat-benchmark number marketed to an agent-workload audience, off by roughly a factor of eight in the setting people actually installed it for. As a project, it's better than its meme: honest about its own measurement problems, and steadily migrating toward input compression, where the real money is. The practical read: install the skill if you *like* terse output — less scrolling, faster reads, no quality loss in lite and full modes (reviewers found ultra mode does drop edge cases). Just don't book it as savings. If you're the one staring at the API invoice, audit your cached-versus-fresh input ratio and your per-turn context size first. That's a duller afternoon than teaching your agent to grunt, and worth an order of magnitude more. ## Sources & further reading - [JuliusBrussee/caveman](https://github.com/JuliusBrussee/caveman)— github.com - [Speaking to AI Agents like Cavemen Saves 65% of Tokens. We Test.](https://blog.jetbrains.com/ai/2026/07/speak-to-ai-agents-like-cavemen-tosave-tokens/)— blog.jetbrains.com - [Caveman Review: The Claude Code Skill That Cuts 65% of Tokens](https://andrew.ooo/posts/caveman-claude-code-skill-token-savings-review/)— andrew.ooo [Priya Nair](https://sourcefeed.dev/u/priya_nair)· AI & Developer Experience Writer Priya covers AI frameworks, developer productivity tooling, and the startup ecosystem across South and Southeast Asia, bringing a researcher's rigour and a practitioner's empathy to every story. She is deeply sceptical of benchmarks and asks hard questions so her readers don't have to. ## Discussion 0 No comments yet Be the first to weigh in.