waitwhat git:20260814.227fdfeA

Immutable. This exact content is served forever at /api/v1/blob/e3b22a3e28715306.

---
name: waitwhat
description: Re-explain what was just said for a reader who got lost — missing premises added, plain language, the project's own vocabulary, and one closing check that the repair landed. User-invoked only; only the reader knows when understanding broke.
disable-model-invocation: true
argument-hint: "optional: the part that lost you"
---

# Wait, What?

The user typing this means one thing: **your map of what they understood was wrong.** An explanation is a checkpoint like any other, and this one failed — the repair is a corrected map, not a louder pitch. Not "be brief", either: an agent told to be brief writes telegrams; an agent told "you lost me" backs up and supplies the ground.

Re-pitch what you just said, for a reader who was not inside your head:

- **Add the missing premises** — the context that makes the conclusion parseable, which you had loaded and they didn't. Behavior level, not implementation level.
- **Speak the project's language** — terms from the project's glossary (`CONTEXT.md`, where it keeps one), never codenames or shorthand this conversation invented. Leaning on a concept no shared vocabulary covers? Define it now — and if there's a glossary, write it in.
- **Shorter and clearer, not shorter and blunter.** Deleting words isn't the job; supplying the ground is.
- **End on evidence, not delivery.** "Explained" is a claim, and its evidence is the reflection, not the speech: close with the one question whose answer shows whether the map is repaired. A wrong answer is a finding about the explanation, not the reader — repair again from where it broke.

A second `/waitwhat` on the same topic is a finding, not a coincidence: the gap isn't the wording, it's a missing shared concept. Name it, define it in the glossary, then explain with it — the repair loop feeds the project's memory, so the next first draft lands on its own.