| name | caveman |
| description | Compress orchestration status into terse, loss-aware bullets for cross-runtime handoffs. Use for low-token status updates, compact worker prompts, compact handoffs, caveman mode, less tokens, terse mode, or compact mode. |
| compatibility | codex,opencode,claude |
| metadata | {"audience":"agents","purpose":"token-compression"} |
Caveman
Respond terse like smart caveman. Keep technical substance exact. Remove filler. This skill affects
agent prose and handoff token volume only; it does not change code, tests, commands, safety gates,
review coverage, or human approval requirements.
Rules
- Drop articles, pleasantries, repeated context, and hedging.
- Keep exact code, commands, file paths, API names, security terms, and failure text.
- Prefer fragments when meaning remains clear.
- Use short labels and direct causality:
problem -> cause -> next.
- Do not hide uncertainty that affects engineering decisions.
- Do not compress irreversible-action warnings, security warnings, approval gates, or multi-step
instructions so much that order becomes ambiguous.
- Never compress exact code, exact commands, exact test output, exact failure text, security
warnings, destructive-action warnings, approval gates, or ordered safety gates.
Agent Use
Use this mode for:
- long-running parent issue orchestration
- worker assignment prompts
- progress updates
- review disposition summaries
- handoff summaries
- repeated GitHub issue/PR status comments
Do not use this mode for:
- code blocks
- exact commands
- exact test output or failure output
- ordered safety gates
- destructive-action warnings
- approval gates
- legal/security disclosures where wording matters
- user-facing product docs unless explicitly requested
Output Shape
Use this pattern where possible:
State: <short state>.
Blocker: <none|specific blocker>.
Next: <one concrete action>.
Evidence: <test/check/PR/issue>.