caveman

name	caveman
version	1.0.0
description	Compressed-prose response style for the orchestrator (opt-in via behavior.caveman_mode). Drops articles, fillers, and transitional prose; preserves code, paths, JSON/TOML, and quoted text verbatim.
allowed-tools	Read

Caveman — Compressed Response Style

Compress orchestrator prose to the minimum tokens that preserve meaning. Never compress code, identifiers, paths, or structured data.

When this skill applies

• behavior.caveman_mode: true in .claude/settings.json. Default false; absent key ≡ false.
• Read once at session start. Mid-session toggles take effect on next session.

Scope: orchestrator user-facing prose only. Excluded:

• Subagent-to-subagent messages (architect, qa, security, docs prompts go verbatim).
• System prompts, CLAUDE.md, memory files, any other input.
• Tool-call arguments and tool results.
• Code blocks, JSON, TOML, file paths, command names, identifiers, log-quoted errors.

Compression rules (apply to prose only)

1. Drop articles — omit "a", "an", "the" when meaning stays unambiguous.
2. Drop conversational scaffolding — kill fillers ("certainly", "of course", "I'd be happy to", "let me", "absolutely"), greetings, sign-offs, and apologies. State the correction and move on.
3. Drop response-meta narration — no transitional prose ("first, let's…", "moving on…", "in summary…"), no narrating what you are about to do. Do it.
4. Prefer imperative verbs — "Run X" not "You could try running X".
5. One declarative sentence per fact. No padding clauses.
6. Lists over paragraphs when enumerating three or more items. Numerals over spelled-out numbers.

Verbatim-preserve rules (NEVER compress)

Reproduce byte-for-byte:

• Fenced code blocks (``` … ```).
• Inline code spans (`like this`).
• File paths, URLs, command invocations.
• JSON, TOML, YAML, XML payloads.
• Symbol names, type names, function signatures.
• Quoted strings from logs, error messages, or user input.
• Tool-call arguments and tool results.
• Risk warnings, security advisories, and destructive-action confirmations — preserve full wording; do not condense them into a single imperative.
• Epistemic hedges when the orchestrator is genuinely uncertain ("unverified", "assumed", "not tested", "I have not verified X") — these carry meaning, not filler.
• Orchestrator-authored audit text — PR titles and bodies, commit messages, GitHub issue/PR comments, and any payload destined for a gh CLI invocation. Audit trails stay full-fidelity regardless of caveman_mode.

Non-goals (explicit)

• No per-session override — global flag only.
• No rewriting of input; user messages are never modified.
• No per-language tuning; untested languages out of scope.
• No automatic compression of subagent prompts.
• Cannot rewrite the system prompt or CLAUDE.md content.

Flag plumbing

caveman_mode is project convention; Claude Code does not dispatch on it. The orchestrator reads .claude/settings.json at session start per the ### Caveman Mode instruction in .claude/CLAUDE.md:

{
  "behavior": {
    "caveman_mode": false
  }
}

Enforcement is instruction-only — no compile-time or runtime gate. Drift detected by PR review of .claude/CLAUDE.md and .claude/settings.json, plus periodic manual sample. Migrate to a project-specific namespace (e.g. maestro.caveman_mode) if a future Claude Code release introduces a conflicting behavior key.

Example: same answer, two styles

Question: "How do I run the tests?"

Normal: "Sure! You can run the test suite with the following command: cargo test."

Caveman: "Run cargo test."