# caveman > A Matt Pocock Claude Code skill that switches the assistant into an ultra-compressed communication mode — drops articles, filler, pleasantries, and hedging while keeping full technical accuracy. Persistent once active, with an auto-clarity exception that suspends compression for security warnings and irreversible-action confirmations. **Use case**: Compress response style until told otherwise **Canonical URL**: https://agentcookbooks.com/skills/caveman/ **Topics**: claude-code, skills, context-engineering **Trigger phrases**: "caveman mode", "talk like caveman", "less tokens", "be brief" **Source**: [Matt Pocock](https://github.com/mattpocock/skills/tree/main/skills/productivity/caveman) **License**: MIT --- ## What it does `caveman` switches Claude into a compressed response style and *stays there* for the rest of the session until explicitly cancelled. The rules: - **Drop:** articles (a/an/the), filler (just/really/basically/actually/simply), pleasantries (sure/certainly/of course/happy to), hedging. - **Use:** sentence fragments, short synonyms (big not extensive, fix not "implement a solution for"), abbreviations (DB/auth/config/req/res/fn/impl), arrows for causality (X → Y). - **Preserve verbatim:** technical terms, code blocks, error quotes. A pattern the skill names directly: `[thing] [action] [reason]. [next step].` Worked example from the SKILL.md: > *Not:* "Sure! I'd be happy to help you with that. The issue you're experiencing is likely caused by..." > > *Yes:* "Bug in auth middleware. Token expiry check use `<` not `<=`. Fix:" Persistence is explicit in the skill: ACTIVE EVERY RESPONSE once triggered. No filler drift, no revert after many turns. Off only on "stop caveman" or "normal mode." **Auto-clarity exception.** The skill suspends compression temporarily for: security warnings, irreversible-action confirmations, multi-step sequences where fragment-order risks misread, and any time the user asks to clarify or repeats a question. Caveman resumes once the clarity-sensitive part is done. ## When to use it Reach for it when: - Token economics matter for the session — long autonomous loops, cost-sensitive workflows, or context-window pressure - The back-and-forth is dense enough that pleasantries add friction the operator notices - You've configured the agent for code-heavy work where prose around the code is overhead When *not* to reach for it: - The session involves walking another operator through a process and friction is the point - You're producing artifacts (blog drafts, customer-facing copy) where the prose itself is the deliverable — compression is for chat, not artifact text - Multi-step destructive sequences where the auto-clarity exception would be triggering on every turn anyway ## Install The skill is distributed via Pocock's [skills repo](https://github.com/mattpocock/skills). Install via his recommended path (`npx skills add` or manual copy of the `SKILL.md` into `.claude/skills/caveman/`) — see the repo README for canonical install instructions. Once installed, trigger with one of the activation phrases. The mode persists until you explicitly cancel. ## What a session looks like The first turn after activation is the calibration: the agent's response gets visibly shorter, technical terms intact, articles and pleasantries gone. After that, every turn maintains the compression unless the auto-clarity exception fires. The persistence rule is load-bearing — without it, the agent tends to drift back to its default conversational tone after a few turns. The "no revert after many turns" instruction is what keeps it pinned. ## Receipts TODO — to be filled in from a real session. **Note on the upstream claim:** the upstream description cites a token reduction figure; under this site's attribution rules we don't republish numerical claims without firsthand verification, so the number is omitted here. When the skill is triggered in production use, capture: actual token-per-response delta over a multi-turn session (measured, not estimated), how often the auto-clarity exception fired, and whether the model drifted back to default style at any point or stayed pinned across the full session. ## Source and attribution Originally written by [Matt Pocock](https://github.com/mattpocock). The canonical `SKILL.md` lives in the [`productivity/caveman` folder](https://github.com/mattpocock/skills/tree/main/skills/productivity/caveman) of his public skills repository. License: MIT. You can install, adapt, and redistribute the skill, with attribution preserved. This page documents the skill from a practitioner's perspective. For the formal spec and any updates, defer to the source repo.