Substance-Safe Output Compressor

The Problem

You tell your AI agent: "Be concise."

It deletes the wrong things.

URLs disappear. Numbers vanish. Reasoning gets cut. What's left is a shorter response — but a worse one.

This skill enforces a two-scope compression model:

Scope	What it covers	Rule
🗑️ Target scope (compress)	Preambles, transitions, filler phrases, repetition, apologies	Always delete
🔒 Protected scope (never touch)	URLs, numbers, code, reasoning, conclusions, named entities	Never delete

Level	Max words	Use case
Lite	15	Light cleanup — remove preambles and duplicates
Full	10	Strong compression — telegraphic style, log format
Ultra	4	Confirmations, status words, yes/no only

Before (verbose):

"That's a great question! I've carefully reviewed your request and I'm happy to help. Based on my analysis, the answer is 42. I hope this helps!"

After (Full compression):

"Answer: 42."

Protected: the number 42. Deleted: everything else.

✅ Deleted:

❌ Never deleted:

Drop SKILL.md into your agent's context or system prompt. The agent will self-select the compression level based on task type.

Auto-selection logic:

Claude Code · Claude API · Any system-prompt-driven agent