Caveman Token Skill
Multi-agent skill/plugin that compresses agent output style to reduce output tokens while preserving technical content. Keep it as a reference/candidate, not an installed default, until the Brin caution/deep-scan result is cleared. Source: JuliusBrussee/caveman, 2026-07-03
What It Is
JuliusBrussee/caveman is a Claude Code skill/plugin with install paths for Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, Copilot, opencode, and other harnesses. It instructs the agent to communicate in compressed prose: fewer articles, fewer polite lead-ins, more direct technical statements. The viral strategy originated from the observation that forcing Claude to "talk like a caveman" can preserve technical accuracy while eliminating verbose scaffolding. Source: X/@om_patel5, 2026-04-06; Source: JuliusBrussee/caveman README and INSTALL.md, 2026-07-03
The screenshot artifact is a real repo snapshot, not just a meme: repo tree includes .agents/plugins, .claude-plugin, benchmarks, caveman, plugins/caveman, and skills/caveman; the sidebar showed v1.1.0 at capture time. Current source has moved further: GitHub API reports 81,611 stars and 4,560 forks, latest release v1.9.0, MIT license, default branch pushed 2026-07-02. Source: local image review, 2026-07-03; GitHub API, 2026-07-03
Relationship to lean-ctx
This complements Lean Ctx by attacking a different side of the token budget:
| Tool | Targets | Mechanism |
|---|---|---|
| lean-ctx | Input context | Trims what the agent reads |
| Caveman skill | Output tokens | Compresses what the agent writes |
Together they squeeze both ends. Treat the exact savings as benchmark-dependent: the X post and README use ~75% framing, while the GitHub repo description says 65%. Even a 40-50% output-token reduction would materially reduce cost and latency on long agent sessions. Source: compiled from X post + GitHub README/API, 2026-07-03
Adoption Status
Do not install this as a default local skill yet. It is high-reputation and appears actively maintained, but the Brin checks on 2026-07-03 returned caution for both the skill surface (score 59, confidence medium, pending deep scan) and repo surface (score 65, confidence medium, pending deep scan). The v1.9.0 release notes add important supply-chain mitigations: install scripts now fetch pinned release artifacts and verify hook files against a committed SHA-256 manifest before writing. That improves the risk posture but does not replace a dedicated local skill audit. Source: Brin API, 2026-07-03; GitHub release v1.9.0, 2026-06-12
Signal Strength
6,292 likes and 6,518 bookmarks (1.04x ratio - unusually balanced, suggesting both entertainment value and practical utility). Source: X/@om_patel5, 2026-04-06
Timeline
- 2026-04-06 | Shared on X by @om_patel5 (6,292 likes, 6,518 bookmarks). Source: X/@om_patel5, 2026-04-06
- 2026-07-03 | Deep artifact review identified the current source as
JuliusBrussee/caveman, with 81,611 stars, MIT license, releasev1.9.0, multi-harness install paths, and Brincautionverdicts pending deep scan. Kept as reference/candidate, not installed. Source: GitHub API, release API, Brin API, 2026-07-03