--- name: operator-rephrase description: "This skill should be used when the last message did not land: say it again in short plain sentences, with no jargon, file paths or issue numbers." --- **Grok Build (`plugins/soleur/lib/harness.ts` `invokeSkill()`):** Read this SKILL.md in this process and run it to completion. A one-segment `soleur:` in this document names a SKILL — on Grok Build, Read `plugins/soleur/skills//SKILL.md` in this process; it is not a nested tool_use. A multi-segment id such as `soleur::` names an AGENT: spawn it, never Read it, and on Grok Build spawn_subagent takes the id with its colons replaced by hyphens (`agentIdToGrokSubagentType`). **Claude Code:** Skill tool for a skill (`soleur:`), Task tool with `subagent_type` for an agent. Forbidden is executing a subset, not the Read. # Operator Rephrase The previous message did not land. Say the **same thing** again, in a register the operator can act on. This is a restatement, not an escalation: no new facts, no extra detail, no lecture. If the content was wrong, fix the content — that is a different job than this one. ## Register As defined in [operator-digest](../operator-digest/SKILL.md) §Register. Read it there; it is not restated here, because a second copy drifts from the first. ### Delta from `operator-digest` Three differences, and only these: - **One message, not a document.** The output is the immediately preceding message said again — not a section, not a report, not a digest. - **Synchronous.** Emit the rephrasing in the current turn. The digest is a scheduled headless artifact; this is a reply. - **The business-consequence rule does not apply.** The digest cuts every line that does not state a business consequence or an owner action. That rule would mangle an in-the-moment restatement of a technical fact the operator asked about. Keep the fact; change only how it reads. ## How to write the restatement The discipline is **influenced by ASD-STE100 Simplified Technical English**. This skill does not claim conformance to that standard — there is no controlled vocabulary in this repository to bind it to, and no gate that could check the claim. What is adopted is the checkable part: 1. Write short sentences. One idea per sentence. 2. Give one instruction per sentence. Do not chain two actions with "and then". 3. Use each word in one sense. Pick the everyday sense and keep it for the whole message. 4. Write in the active voice. Name who does the thing. 5. Remove jargon, file paths, issue numbers, PR numbers, and command names. Name the effect instead of the artifact that produces it. **Exception: keep any literal the operator must type, click or paste** — a command to run, a button label, a value to enter. Put it alone on its own line, unchanged; the sentences around it say what it does. A restatement that deletes the one string the operator needs has removed the action along with the jargon. 6. Open with the outcome. Put the reason after it, if the reason is still needed. ## Vocabulary **The glossary is a stop-list, not a source of approved terms.** Rule 3 above is still satisfied by choosing plain everyday words, not by looking a term up — there is no controlled vocabulary of approved founder words here and there is not meant to be one. What `knowledge-base/project/glossary.md` gives this skill is the inverse, and it serves **rule 5**: it is the list of internal words that **must** be translated before the operator sees them. Each entry — or each sense of a multi-sense entry — opens with a plain clause saying what the thing is, and that clause is what you translate *into*. The repo-relative path in the entry body is for the agent; it is one of the things rule 5 strips, so it never appears in the restatement. Use it as a check, not as a lookup table: if a word you are about to say has an entry, say the plain clause instead of the word. A word with no entry is not thereby approved — plain language is still your own judgment. Do not point at `plugins/soleur/docs/pages/glossary.njk` — that is marketing surface, its entries are category terms rather than operational ones, and it is a Nunjucks template rather than readable plain text. `knowledge-base/marketing/brand-guide.md` is the founder-facing lexicon and is the right place for a word the operator *should* see. ## When to self-invoke Invoke this skill in the same turn, without being asked by name, when the operator's reply is one of these shapes — each is "that did not land", not a new question: - "what?", "huh?", "sorry?" - "in plain English", "in normal words", "without the jargon" - "I don't follow", "I don't understand", "lost me" - "say that again", "come again", "one more time" - "what does that mean for me?", "so what do I do?" Do **not** invoke it when the operator asks for *more* detail, disputes a fact, or asks a follow-up question — those are new answers, not restatements. ## Stop conditions - Stop after one restatement. If that one also does not land, ask which part did not land rather than rephrasing a third time. - Do not apologise at length. One short acknowledgement at most, then the restatement. - Do not change the decision, the recommendation, or the facts while rephrasing them. A restatement that says something different is a new answer wearing a rephrasing's clothes. ## Siblings `soleur:operator-digest` is the scheduled weekly counterpart: same register, batched, written to a file rather than said in a turn.