caveman

name: caveman description: >- One-response ultra-compressed communication overlay for casual Gandalf-to-user prose. Use when the user says "caveman", invokes /caveman, or asks for brief plain-prose output with cues like "be brief" or "less tokens".

Use this skill as a single-response compression overlay. It is not a workflow gate, not persistent state, and never carries to the next turn.

Scope: Gandalf-to-user casual prose only. Do not apply to tool calls, delegated prompts, artifacts meant for paste, or any response where fidelity matters more than token savings.

Triggering

Explicit caveman or /caveman always triggers for the current response.
Soft cues such as be brief, less tokens, or short answer trigger only on plain-prose turns.
Soft cues never override the never-apply list below.
There is no sticky mode. A later response must be triggered again.

Compression style

Target roughly 75% reduction while keeping technical meaning exact:

Drop articles, filler, pleasantries, and hedging.
Use fragments when clear.
Abbreviate common technology terms: config, auth, DB, req, res, fn, impl.
Use arrows for causality or flow: X -> Y.
Prefer one direct sentence over explanation when enough.

Pattern: Thing -> cause/effect. Fix/next step.

Hard never-compress verbatim

Keep these exactly as written:

Code blocks.
Error messages.
Exact identifiers, paths, and commands.
Structured data: JSON, XML, YAML, tool-call envelopes, tables, and any machine-parsed output.

Hard never-apply

If any of these are present, the entire response uses full fidelity. Do not mix caveman and normal style in one response.

Delegation or subagent prompts.
Plans.
Council or Saruman verdicts.
Approval-gate text.
Security warnings.
Irreversible-action confirmations.
Paste or commit deliverables: handoff docs, commit messages, PR bodies, issue bodies, and similar artifacts.

Tie-break

If unsure, use full fidelity. If a never-apply condition is detected mid-response, the entire response yields to full fidelity.