machine-voicelisted

# Machine Voice Compression for output whose next reader is scanning for information. ## Invariant **ALWAYS** confirm the element matches the list below before compressing. **NEVER** apply these layers to prose a human reads — that is `human-voice`'s. ## The list This skill governs an output element when it is one of: - Agent trace or tool-call log - Status line, progress update, handoff note - State dump, config, schema, reference card - Structured data block another agent will parse **These four bullets are the whole list.** Do not generalize it. Not on the list → this skill does not apply; follow `human-voice` for that element. The split is **per element, not per response**. A reply that explains something in prose and embeds a trace uses `human-voice` for the explanation and this skill for the trace. ## Precedence — read before applying the list The list above is not the last word. Two rules beat it, in this order. **1. Out of scope beats the list.** These never get compressed, even when they look exactly like a config or a schema: > Code and file contents, commit messages, creative or persona writing, and > turns that are only a clarifying question or only tool calls. The contents of a real `.eslintrc`, a `SKILL.md` being written, or a JSON Schema the user asked you to author are **file contents**. Ship them verbatim and unstyled. Compressing a file the user will save corrupts it — Layer 1 rewrites content, and content is the artifact here. **2. Serving pro