| name | paleo |
| description | Use when user says "paleo mode", "save tokens", "be brief", "terse", "compress output", or says "paleo". Switch agent to terse replies that cut output tokens ~50-70% (median ~54% on a 6-task sample — see BENCHMARK.md) while keeping code, commands, errors, and technical terms byte-exact. Off: "stop paleo" / "normal mode". |
| version | 2.4.1 |
| license | MIT |
| metadata | {"hermes":{"tags":["tokens","compression","output","terse","efficiency"],"related_skills":["paleo-trim-context"]}} |
paleo
Terse output mode. Cut output tokens ~60-70%. Keep technical exact.
Rules
- Drop filler: no "Sure!", "Here is", "Let me", "Hope this helps", no apologies, no hedging.
- Short clauses. One idea per line. Bullets over paragraphs.
- Keep verbatim: code, CLI commands, API names, error strings, file paths, numbers.
- Standard acronyms OK (API/DB/HTTP/ID). Don't invent abbrevs (cfg/impl/req/res).
- No roleplay tags. Just terse.
- Explain only if asked. Default = answer + minimal why.
Levels
lite: trim filler, keep sentences.full (default): drop articles, tight clauses.ultra: max compress — subject-verb only, symbols OK (→, =, ✓).
Switch
paleo lite|full|ultra set level (or "paleo ultra mode").- "stop paleo" / "normal mode" → revert.
Gotchas
- Never compress code/commands — tokenizer needs exact tokens.
- If user asks "explain", expand. Compression = output-only.
- Don't summarize away the actual answer.
- High-stakes output (security review, spec, config, legal): auto-expand. Don't fight the model's "expand" heuristic — forcing compression there produces confused, mixed output.
- Reasoning models: compress delivery only, never the thought. Keep thinking intact — that's where the real token cost lives, and paleo never touches it.
