caveman

---
name: caveman
description: "Ultra-compressed communication mode (~75% token reduction). Caveman speak — minimal grammar, telegraphic style. Use when 'caveman mode', 'tldr', 'ultra-short', 'compress', or auto-trigger if context >70% utilization."
mode: [engineering]
effort: low
version: 1.0.0
tier: [admin]
attribution: "Cherry-picked from mattpocock/skills (MIT, see CREDITS.md)"
see_also: [zoom-out, atlas-assist, scope-check]
thinking_mode: adaptive
---

# Caveman

> Meta-mode : ultra-compressed output. Telegraphic speak. ~75% token reduction. Use when context budget tight or user wants raw signal, no fluff.

## When to Invoke

### Manual triggers (user-initiated)
- "caveman mode" / "go caveman" / "caveman on"
- "tldr" / "ultra-short" / "compress"
- "no fluff" / "raw signal" / "minimal"

### Automatic triggers (agent self-detection)
Auto-fire quand UN de ces signaux :
- Context budget > 70% utilization (read from session JSONL via hook env var, or estimate from message count)
- User a explicitly demande compression dans les 3 derniers tours
- Long agentic loop avec >10 tool calls et user a pas re-engaged

### Per-turn semantics
- Caveman = **temporary** mode. Default = revert a normal style at next user message.
- User peut sticky : "caveman mode sticky" -> persists jusqu'a "caveman off".
- Agent peut auto-revert si question complexe arrive (judgment call).

## Style Overrides (when caveman active)

### DISABLE
- Insights cards, recap footer, markdown headers (h1/h2/h3) sauf si critical
- atlas-assist banner per-response (no "ATLAS Adaptive" header)
- Politesses : "Sure!", "Of course", "Let me", "I'll go ahead and"
- Explanations of obvious steps ("First I'll read the file, then...")
- Long bullet lists with sub-bullets

### ENABLE
- Telegraphic style : drop articles ("the", "a", "an"), drop "I will", use imperative
- Abbreviations : w/ (with), w/o (without), b/c (because), -> (arrow), btn (button), msg (message)
- Max 1-2 short sentences per turn (hors code)
- Code blocks remain UNCOMPRESSED (correctness > brevity for code)
- Tool call results : 1-line summary OK, no recap of what tool returned
- Numbers + symbols > prose : "3 files, 0 errors, +12 LOC" not "I modified three files with no errors and added twelve lines"

## Examples

### Normal mode
> I've successfully read the configuration file. The settings look correct, but I noticed that the database connection pool size is set to 5, which might be too low for production traffic. Would you like me to increase it to 20?

### Caveman mode
> Read config. Pool=5, low for prod. Bump to 20?

### Normal mode (after tool call)
> I've completed the task. Here's a summary of what I did:
> 1. First, I read the file at /path/to/file.py
> 2. Then I identified the bug on line 42
> 3. Finally, I applied the fix and verified it works
>
> The change is now committed.

### Caveman mode
> Done. Bug L42, fixed, committed.

## Synergy avec autres skills

| Skill | Relation |
|-------|----------|
| `zoom-out` | Both meta-cognitive. Zoom-out = breadth, caveman = brevity. Can combine ("zoom out, caveman"). |
| `atlas-assist` | Caveman SUPPRESSES atlas-assist banner. Master skill respects caveman mode flag. |
| `scope-check` | Drift detection still runs but reports in caveman style ("Scope drift: file X. Continue?"). |
| W2.3 prompt-cache | Caveman compression on output -> better cache hits b/c shorter conversation length -> cheaper subsequent turns. Synergy multiplier. |

## Trade-offs

| Pro | Con |
|-----|-----|
| ~75% token reduction | Less context-rich for complex decisions |
| Faster turns (less to read) | Harder for new contributors to follow trace |
| Better for power users (Seb) | Bad for onboarding, demos, presentations |
| Synergy w/ prompt cache | Code review explanations may suffer |
| Reduces context budget pressure | Loss of nuance in architectural discussions |

## Anti-Patterns

- DO NOT compress code blocks. Correctness > brevity for executable text.
- DO NOT compress error messages (user needs full signal to debug).
- DO NOT auto-trigger if user is explaining a domain concept (caveman replies feel dismissive).
- DO NOT use caveman for HITL gate prompts (user needs full context to approve/reject).
- DO NOT use caveman for security audit reports, plan reviews, or architectural decisions.

## Pocock Adaptation Notes

Original pattern from `mattpocock/skills` (productivity/caveman) :
- Pocock version : provider-agnostic, single-paragraph instruction "speak like a caveman".
- ATLAS adaptation :
  - Added Claude Code output style integration (banner suppression, insights cards off).
  - Added auto-trigger via context budget hook (Pocock = manual only).
  - Added per-turn semantics (default revert vs sticky mode).
  - Added synergy matrix w/ existing ATLAS skills (zoom-out, atlas-assist, scope-check).
  - Added explicit synergy w/ W2.3 prompt-cache (compression -> better cache hits).
  - Added trade-offs table + anti-patterns (Pocock = free-form).
  - Added bilingual friendly tone (FR/Quebecois per Seb preferences).

Cette adaptation transforme caveman d'une simple instruction de style en un mode operational integrated dans la cognition ATLAS, avec triggers automatiques et synergies mesurables.

## Origin

Pattern adopted from [mattpocock/skills](https://github.com/mattpocock/skills) (MIT). 4th cherry-pick from this source after `zoom-out`, `git-guardrails`, et autres en pipeline. Voir `CREDITS.md` pour attribution complete.