paleo

What paleo does

paleo is a set of seven token-saving skills for LLM agents that trims output and context without changing your code. The base skill, paleo, puts the agent in a terse output mode that drops filler while keeping code, commands and technical terms byte-exact. The others add automatic detection (paleo-auto), a hard per-task token budget (paleo-budget), proactive context trimming (paleo-trim-context), condensing of older chat turns (paleo-converse), tight summaries of bulky tool output (paleo-summary) and compact JSON output (paleo-json).

Skills are triggered by plain English phrases such as be brief, budget 2000 or trim context, so no slash commands need registering. Each SKILL.md is deliberately short to keep loading cheap. The README reports about 50 to 70 percent fewer output tokens, with a median of roughly 54 percent on a six-task sample documented in BENCHMARK.md. It follows the open Agent Skills standard and names Claude Code, Codex, Gemini CLI, Qwen Code, OpenCode, Cursor, GitHub Copilot, Cline and Windsurf, with 40-plus agents reachable through npx skills add. The repository description lists Hermes among them.

Key features

• Terse output mode that keeps code, commands and technical terms byte-exact
• paleo-auto, which watches the session and enables the right skills automatically
• paleo-budget, a hard per-task token cap that summarizes the tail
• Context trimming and chat condensing with the last N turns kept verbatim
• Summaries of bulky tool output and compact JSON output
• Plain English trigger phrases, with each SKILL.md kept short

When to use it

• Lowering token cost in long, chatty Hermes sessions
• Setting a hard budget such as 2000 tokens for a single task
• Condensing large logs or diffs before they fill the context window

Who it is for: Agent users who pay per token or hit context limits and want drop-in skills that shorten output and context.

How it fits with Hermes Agent

A general skill pack that lists Hermes among the agents it supports through the Agent Skills standard.

FAQ