Caveman Audited: What 3% Skill and 33.2% Proxy Cuts Actually Mean
An audit of JuliusBrussee/caveman using its own evals, SoloSoft's audit, and JetBrains data—separating the 3% voice skill from the 33.2% local proxy.

Q. Does the Caveman skill really cut AI coding agent tokens by 65%? No. Turning on the default
/cavemanvoice skill alone only saves a median 3% (-2.88%) of output tokens compared with a plain"Answer concisely."prompt. The meaningful cost reduction comes from the local proxy (@caveman-ai/cli), which shrinks logs, JSON, CSV, and test outputs before the agent reads them, cutting full-session input tokens by 33.2%.
⚠️ Accuracy guardrail: This post reflects information verified as of November 19, 2026. Star counts, benchmarks, and CLI flags may change as the repository evolves. Because the original JetBrains blog post (
blog.jetbrains.com) returned HTTP 403 and could not be fetched directly, JetBrains figures (86 coding tasks, 8.5% fewer output tokens, p=0.82) are cited only as reported in theJuliusBrussee/cavemanREADME and the SoloSoft audit.
30-second summary
- Where the 65% headline came from: It traces back to older README estimates and a retired stats constant; in the repository's own committed eval (
claude-opus-5-5, 10 dev questions), default/cavemancuts 38.3% versus no prompt at all, but only 3% at the median versus"Answer concisely.". - Where the real savings happen: Rather than the output voice skill (Say less), the local proxy (Read less,
@caveman-ai/cli) that compresses tool output cut total input tokens by 33.2% (885,793 to 591,673) across six large-output workloads. - Where it backfires: The skill instructions add roughly 1,160 input tokens of prompt overhead, and HTML pages—which currently have no compressor—ended up 9.9% larger in token count through the proxy.
- Who benefits most: For pay-per-token Claude Code or Codex users hitting context limits, pairing the local proxy and
CLAUDE.mdcompression (/caveman-compress) delivers far more value than installing the voice skill alone.

What Caveman is and why it crossed 120K GitHub stars
Caveman (JuliusBrussee/caveman) is an open-source skill and local proxy that makes coding agents say less (Say less) and read less (Read less). Released in April 2026 under the Apache-2.0 license, it reached 121,362 GitHub stars and 7,097 forks as of November 19, 2026.
The core idea borrows from ASD-STE100, the controlled English used in aircraft maintenance manuals. It strips greetings, filler, and recaps, forcing replies into a [thing] [action] [reason] pattern while keeping code blocks, CLI commands, file paths, exact error strings, and negations (not, never, no) verbatim.
Starting in v4.0, Caveman expanded beyond a prompt skill into six modular parts ("Six rocks") bundled inside @caveman-ai/cli: the voice skill, tool-output proxy, difficulty-based model routing (Auto), reusable script execution (Blocks), prompt-cache warming (Waste fixes), and web page compression (Browse). As covered in Toris Studio's harness engineering post, the surrounding harness often determines real-world cost and reliability more than the base model alone.
Say less: why 65% turns into 3%
The default /caveman skill cuts 38.3% of output tokens against an unprompted verbose agent, but only a median 3% (-2.88%) against an agent already told "Answer concisely." Modern coding models already know how to be brief when asked in one line.
SoloSoft's October 3, 2026 audit independently recomputed the committed evals/snapshots/results.json snapshot (claude-opus-5-5, 10 dev questions, tiktoken o200k_base) and confirmed the gap.
| Condition | Output tokens | Median vs terse control | Notes |
|---|---|---|---|
| No system prompt (baseline) | 6,983 | — | Includes verbose explanations |
"Answer concisely." |
4,334 | 0% (control) | Practical everyday baseline |
/caveman (default) |
4,119 | -2.88% (~3%) | Adds ~1,160 rule input tokens |
/megacave (Classical Chinese) |
3,683 | -9.14% | Character drop outpaces token drop |
/ultracave (max compression) |
2,693 | -34.73% (~35%) | Fragments and keywords only |
If you pay per request rather than per token (such as GitHub Copilot premium requests), shorter answers save nothing on your bill. Worse, loading ~1,160 tokens of skill rules into the prompt can increase input cost on short Q&A turns depending on prompt caching.
Read less and memory compression: where real savings come from
Most tokens in a coding-agent session are not the agent's prose—they are tool outputs flooding the input context. Reading test logs, CSVs, YAML configs, and JSON payloads repeatedly burns tens of thousands of tokens per turn.
According to the official docs (docs.caveman.so) and repository README benchmark (six synthetic large-tool-output tasks, 18 pairs, August 2026 engine), compressing what the agent reads moves total session cost by an order of magnitude more than changing its tone.
| File type | Single file tokens | File saved | Full-session input tokens | Session saved |
|---|---|---|---|---|
| CSV | 28,041 → 314 | 98.9% | 165,823 → 74,484 | 55.1% |
| Logs | 22,810 → 348 | 98.5% | 148,807 → 74,068 | 50.2% |
| YAML | 20,447 → 178 | 99.1% | 132,124 → 71,027 | 46.2% |
| Test output | 18,806 → 203 | 98.9% | 150,377 → 108,514 | 27.8% |
| JSON | 18,837 → 281 | 98.5% | 147,975 → 108,939 | 26.4% |
| HTML (unsupported) | 21,670 → 21,670 | 0% | 140,687 → 154,641 | +9.9% worse |
| All six combined | 130,611 → 22,994 | 82.4% | 885,793 → 591,673 | 33.2% |
All 18 tasks produced the exact right answer (18/18), and every original file stays on disk so the agent can retrieve full lines whenever needed. HTML had no compressor in that suite, so Caveman paid its overhead without saving input tokens.
The companion /caveman-compress skill shrinks natural-language prose inside CLAUDE.md and project memory files loaded on every session start. Across five shipped fixtures, it cut 6,198 tokens down to 4,138 (33.2% saved) while leaving headings, code blocks, URLs, and file paths untouched.
Q. Should I also write my own prompts in caveman style to save more tokens? No. Both the official docs and Adobe Research's CAVEWOMAN paper cited by the repo warn that compressing the user's own prompt degrades answer quality and triggers follow-up turns.
Installation and telemetry caveats
Installing only the voice skill takes one npx skills command, while the input-compressing proxy requires @caveman-ai/cli on Node.js 22.13 or newer. It fits directly into the plugin and extension workflow discussed in our Claude Code workflow post.
# 1) Voice skill only (Claude Code, Codex, Cursor, Gemini CLI, etc.)
npx skills add JuliusBrussee/caveman -g
# 2) Full setup including the local proxy (requires Node.js 22.13+)
npx @caveman-ai/cli@latest
One operational detail to check before installing the CLI: while the standalone skill runs locally and sends nothing, the caveman CLI sends telemetry by default (commands run, token counts, a random install ID, OS, and IP). It states that prompts, code, and file paths are never collected, but if your workplace policy forbids external telemetry, turn it off immediately.
caveman telemetry off
# Or set the environment variable
export DO_NOT_TRACK=1
Toris's take
In production solo-dev workflows, prioritize the local proxy (@caveman-ai/cli) and CLAUDE.md compression (/caveman-compress) over using the voice skill alone.
- Skip it if: You pay per request (like GitHub Copilot premium requests) or already enforce concise replies in your system prompt—the default
/cavemanvoice only saves about 3% there. - Adopt it if: You run Claude Code or Codex against verbose test logs, server logs, and large JSON payloads; the local proxy cuts 26–55% of input tokens and directly helps with context limits and API bills.
- First thing after install: Run
caveman telemetry offon work repositories, and disable modules that do not fit your workload (for example, HTML-heavy scraping) withcaveman off.
Sources
Advertisement