--- name: rem-muse description: Think one page through slowly over several turns in one thread — ask many questions, pick the five worth pursuing, study each, and write only when a question yields an insight. --- # Muse on one page Input: a written page, its cited originals, the open leads from enrich, the questions earlier runs left, and the pages it links to. Follow the page kind's `rem-page-*` rules and `rem-investigate`'s quote and evidence rules unchanged. Contact details keep the label enrich gave them; a number the subject publishes on their own site stays theirs, `[personal]`, even when a company page lists it too. Muse is thinking, not searching. Everything you need is already in the originals; what is missing is the right question. **Asking a precise question matters more than answering it.** Go slowly. One sharp insight is worth more than ten new facts. Muse runs as one conversation. Each turn does one step below, then stops and waits for the next turn; what you read and thought in earlier turns stays with you, so do not reread what you already read. ## Turn 1 — ask, then choose Read the page, then every cited original whole. Do not edit the page. Write at least twenty questions this material raises. A good question: - is **specific**: names the person, the date, the number, the source ("Why did replies stop after the 2026-08-20 price change [6]?"), not "what is the relationship like?"; - is **answerable** from originals, or says exactly what record would answer it; - **matters to the user**: its answer would change what the user does next, what they believe about this person or project, or how they deal with them. Ask from several angles: - **Stakes**: what does the user want from this, and they from the user? What is lost if nothing happens? - **Change**: what is different now from the first contact? When did it turn, and what happened just before? - **Pattern**: what keeps recurring — who chases, who delays, what topics, what conditions? Is it really a pattern (three or more cases) or one incident? - **Contradiction**: where do two sources, or the page and a source, disagree? What did someone say and then not do? - **Silence**: what was asked and never answered? What should have happened by now and did not? - **Connection**: who else is in these threads, which project or organisation is this really about, which other page holds the other half? - **Why**: why did the user decide what they decided; what principle shows? - **The page itself**: which sentence would mislead the user if they acted on it today? Which open thread is actually closed? Is another page the same person or project (same name, an address or handle that the material ties to them)? Is each current `Insight` bullet, and the lead, still true and still the most important thing — or left over from an earlier, thinner page? Then **choose the five** most worth pursuing: the ones whose answer would most change what the user understands or does, and that the originals can plausibly settle. - If another page may be the same subject, that question is chosen first. - At least three of the five are about the subject and the user — stakes, change, pattern, silence — not about one document's details. - No two chosen questions come from the same thread. Reply with all questions, the five marked and why each was chosen: ``` * 1. — why it matters * 2. ... 6. ... ``` ## Turns 2–6 — study one question each Each turn takes the next chosen question. Reread the originals that bear on it, search the notebook's archives if a record you need is named there, and think. Decide one of: - **insight** — the answer shows something the user could not see from any single message and that changes what they believe or do: a stake, a turn, a pattern, a contradiction, a thread that is really closed, a sentence on the page that is wrong. Test it before you call it one: if a single message says it, or the page already says it, it is not an insight. Your reply names the two or more sources it needed. **Only now edit the page**: put it, cited, in the right place (`Insight` as `At stake:`/`Changed:`/`Pattern:`, a turn in `History`, a closed thread out of `Open threads`, a correction where the error was). If it changes who they are, what is live or what the user does next, rewrite that sentence of the lead now. - **no insight** — the answer is a restatement, a detail, or already on the page. **Do not edit the page.** Say so; that is a normal result. A page that already *links* another page has not said they are the same person or project; confirming that from the originals is an **insight** — say it on the page with the tie (shared address, signature, name), and fill this page's `Unknown` facts that the other subject's originals settle. These are never insights: that one record does not prove another; that more evidence is needed; that two documents describe different things; the same fact with more quotes. If you can only say it as "Inference: … does not establish …", it is **no insight** or **open**. - **open** — the originals cannot settle it. Do not edit the page. Name the record that would answer it; it becomes enrich's next lead, and if it matters, the last turn puts it in `Uncertainties` as the question. Never answer from plausibility. An inference is labelled and rests on quotes. Each edit is small and cited; never rewrite a correct section, never drop a cited fact that is still true, keep every quote verbatim. An `Insight` bullet states what is true and why it matters — a finding, not advice and not a hedge. It never tells the user what to do (`verify`, `agree`, `complete`, `confirm`, `pursue`, `should`): the action goes in the lead's next step or in `Open threads`. Not `At stake: confirm the renewal before 30 November`, but `At stake: the contract renews at A$15k on 2026-11-30 unless one side gives notice, and neither has [4][9]`. It never starts with or contains `Inference:` — if it needs that label it is not ready. `Insight` keeps at most four bullets, so a new one replaces the weakest (move that one's fact, quotes and citations where they belong). An insight adds a sentence or two, not a paragraph of supporting quotes. The owner is `the user`, never their name. Reply each turn: ``` verdict: insight | no insight | open ``` ## Last turn — look back Only if the page was edited this run: reread it top to bottom and remove what the new insights made superseded or said twice. Then check the lead against every insight this run added and every `Insight` bullet. Ask of each: does it change who they are to the user, what is live, or what the user does next? If yes, the lead must say so — rewrite its sentences to name the most consequential live item and the user's actual next step as they stand now, keeping each quote the lead had (move one you no longer lead with to where it belongs). A lead that still reads as the page did before this run, after an insight changed what is live, is wrong. The lead stays the page rule's three short sentences with a few citations each: it names the one live item and the next step, not every finding. The live item is the one with the most at stake for the user — what they owe, money, a deadline, a decision only they can make — not the newest message; an agent's status report is never the ask it quotes. A same-person finding goes in `Facts` and `Insight`, not into the lead. A project page has no lead: its first lines of `Where it stands` and `Open threads` carry what is live and next, and `What it is` stays the product description. In your reply, give the lead before and after, or the insight that left it standing and why. The page should leave this run no longer than it came in, apart from the insights. Then bring the page back to its kind's shape, which enrich does not do: `History` (a project's `Key decisions`) at most sixteen lines — fold the oldest lines into one dated sentence per period (`- 2026-07 to 2026-09: [n][n]`), never a semicolon list; `Our relationship` a state, with long contract quotes cut to the clause the user acts on; each `Uncertainties` bullet one question ending in `?`, then its citations; every `Insight` bullet a finding with no instruction in it; no blank lines inside a list. Removing repetition merges sentences; it never removes a `“…”` quote or a citation — compare with the page at the start of the run and restore any that went missing. Every `[n]` must be defined under `Sources`. Reply: insights added, questions left open (ranked, with the record that would answer each), and pages that should be linked or merged — never merge. The next run starts from these open questions; ask better ones than this run.