# Huh > When a message did not land or pasted agent output needs translation, restate it plainly, expand shorthand, separate results from requests, and flag unsupported claims. - Skill: `timharris707/huh-2` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add timharris707/huh-2` - Raw SKILL.md: https://api.skillmd.com/api/skills/timharris707/huh-2/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: timharris707 (https://skillmd.com/u/timharris707) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/timharris707/huh-2 --- # Huh Read the [Codex desktop binding](../../../CODEX.md) when this workflow needs harness mechanics, model routing, or recovery. Decode one message that failed its reader. The explanation translates the target and adds no unsupported facts. If this is a side question during authorized work, answer it first and then continue that work unless the user paused or canceled it. The explanation itself authorizes no new changes. ## Target Pasted text: decode that. Nothing pasted: decode your own most recent substantive message. Skip back past courtesies to the one carrying content, and use the conversation it sits in to resolve its references. A target from another session arrives without its context: resolve only what the text itself supports, and mark the rest unknown rather than guessing. ## The four moves Work every move on the target. 1. **Restate.** Rewrite the whole message in complete plain sentences at the reader's altitude, defining every term of art in the sentence that first uses it. 2. **Expand the shorthand.** List every codename, label, abbreviation, or numbering scheme the message coined, each mapped to plain words, or marked unknown when the target does not say. 3. **Split report from request.** State separately what has already happened and what the reader is being asked to decide or do. When nothing is asked, say so outright; a reader hunting for a hidden question is half the confusion. 4. **Flag the unshown.** Name each claim presented as settled that the message backs with no evidence, number, or verification. This is bookkeeping, not accusation: "stated, not shown." ## Output shape The restatement leads. After it, only the moves that found something, each a short labeled section: **Shorthand expanded**, **What's being asked of you**, **Stated but not shown**. The whole answer runs shorter than the target unless the target was too compressed to understand. ## Done when (checkable: verify each line before reporting complete) - Every term of art in the restatement, whether coined by the target or inherited jargon, is defined in the sentence that first uses it. - Every coined term is expanded or marked unknown; none is silently dropped. - The reader can quote what is being asked of them, or the answer says nothing is. - Every settled-sounding claim either carried its evidence in the target or is flagged. - The explanation added no unsupported fact or new scope; any ongoing work respected the latest user instructions. ## Attribution The trigger is the same human moment as Matt Pocock's [wait-what](https://github.com/mattpocock/skills/tree/main/skills/productivity/wait-what) (MIT): the reader stops the agent and asks it to land the message this time. His skill is a two-line re-pitch prompt; the four moves, the target rules, and the output shape here were written independently for this catalog, from failure modes observed in its own sessions.