Update to Tabularia

{
  "description": "Read a Tabularia chapter's prose at the sentence level — rhythm, repetition, adverbs, dialogue balance, readability — from counts Tabularia measures rather than numbers you estimate. Use when the Writer asks how their prose reads, whether it is repetitive, whether the sentences are samey, whether there is too much dialogue or description, whether the point of view slips, whether the characters sound alike, or asks for a style or readability check.",
  "included_files": [],
  "name": "prose-read",
  "skill_md_contents": "---\nname: prose-read\ndescription: Read a Tabularia chapter's prose at the sentence level — rhythm, repetition, adverbs, dialogue balance, readability — from counts Tabularia measures rather than numbers you estimate. Use when the Writer asks how their prose reads, whether it is repetitive, whether the sentences are samey, whether there is too much dialogue or description, whether the point of view slips, whether the characters sound alike, or asks for a style or readability check.\n---\n\n<what-this-is>\n\nThe half of a style report a model is actually good at.\n\nTabularia counts: sentence lengths and their spread, words echoing close to\nthemselves, the content words used most, adverbs per thousand words, how much\nof the passage is spoken aloud, and the two Flesch readability figures. You read\nthose counts and say which of them matters in **this** chapter, for **this**\nWriter.\n\nThe counts work for any prose: a novel, a nonfiction book, an article. Read\nthem against the form. Dialogue ratio means little in an essay, and a\nreadability grade matters more in an article read on a phone than in a novel.\nA screenplay is not prose; use `screenplay-read`.\n\n</what-this-is>\n\n<never-estimate-a-number>\n\nThis is the rule the whole skill rests on.\n\n`tabularia_prose_metrics` returns counts measured in Tabularia. Every figure it\ncovers must come from it. Do not estimate a word count, do not work out an\naverage sentence length, do not say \"roughly a third of this is dialogue\" from\nreading it, and do not invent a score of your own from its numbers. Readability\nis already there: `readability.fleschKincaidGrade` and\n`readability.fleschReadingEase`, from a syllable rule that is close rather than\nexact, so quote them as approximate.\n\nA number guessed by eye is the problem, not a number as such. You will be\nconfident and you will be wrong, and a Writer has no way to check. A number you\ninvented is worse than no number, because it looks like the ones that are real.\n\nSo when the Writer asks for a count the tool does not measure — passive\nconstructions, filter words, dialogue tags, sentences opening with the same\nword — give it to them the first time they ask, and make it one they can check:\n\n1. Read the whole passage the count covers, with `tabularia_read_scope` or\n   `tabularia_read_window`, not a sample of it.\n2. Find every instance, and list them with their Block indexes. The list is the\n   count; the number is only its length. If you can run code, count with it over\n   the text you read.\n3. Say plainly that this count is yours, not Tabularia's, and say what you\n   counted as an instance — \"was\" plus a past participle, say — because the\n   Writer's idea of passive voice may not be yours.\n\nA long passage with many instances: give the total, the list for the first\nstretch, and offer the rest, rather than refusing or rounding.\n\nDo not volunteer unmeasured counts in an ordinary read. They belong to the\nWriter's question, not to your report.\n\n</never-estimate-a-number>\n\n<how-to-read>\n\n1. `tabularia_list_scopes`, then `tabularia_prose_metrics` with `scope: \"chapter\"`\n   or `\"scene\"`. A whole-Manuscript read averages away the thing you are looking\n   for: one slow chapter disappears into eighty that are fine.\n2. Run it again on a chapter the Writer is happy with. **A number means nothing\n   on its own.** Fourteen adverbs per thousand words is neither good nor bad; it\n   is only interesting if their other chapters run at five.\n3. `tabularia_read_scope` to read the passage itself. The counts tell you where\n   to look; they never tell you what is wrong.\n\nCheck `tabularia_list_drafts` first — the counts measure what is showing.\n\n</how-to-read>\n\n<what-the-numbers-are-worth>\n\n- **Sentence spread** is the one that usually matters. Rhythm is variance, and\n  `standardDeviation` near zero over a long passage is the thing a reader feels\n  as monotony even at a comfortable average. Read `lengthBuckets`, not the mean:\n  forty words every time and a mix averaging forty read nothing alike.\n- **Echoes** are sorted by distance. A distinctive word twice in ten words is\n  worth a look; the same word twice in fifty usually is not. Read the paragraph\n  before mentioning it — repetition is often the point.\n- **Frequent words** show a Writer their own tics. Say what you notice; do not\n  hand over the list. Twenty rows is data, three observations is a note.\n- **Adverbs** are a habit, not an error. Never advise removing adverbs as a\n  rule. If the rate is high _for this Writer_, name two or three where the verb\n  was already doing the work.\n- **Dialogue ratio** is absent when the passage uses no quotation marks, which\n  means dashes or none — not silence. Never report zero as a finding.\n- **Readability** is a school-grade formula built for textbooks, and fiction\n  runs low on it by design: short sentences and plain words are a style, not a\n  fault. Compare it across the Writer's own chapters, the same as everything\n  else, and never present a grade as a target.\n\n</what-the-numbers-are-worth>\n\n<what-the-counts-do-not-see>\n\nEditors name a handful of problems more often than any other, and none of them\nis a count. Read for them when the Writer asks how a chapter reads, and raise\none only when it is plainly there:\n\n- **Point of view that slips.** In a scene held in one character's head, a\n  sentence that knows what someone else is thinking or feeling. Ask of each\n  paragraph whose head it is in. A change of viewpoint at a scene break is a\n  choice; one in the middle of a paragraph is the slip. Omniscient narration\n  is not head-hopping, so read a few pages before deciding which this book is.\n- **Characters who all sound alike.** Cover the tags and see whether you can\n  tell who is speaking. Point to two characters whose lines could be swapped.\n- **Dialogue with nothing around it.** A page of lines with no action, setting\n  or thought between them, where the reader loses who is where.\n- **Characters naming each other.** People in conversation rarely say each\n  other's names; a scene where every third line does reads as stiff.\n- **Exposition in the dialogue.** One character telling another what they both\n  already know, for the reader's benefit.\n\nCite the Blocks for each, like every other finding. Two of these in a read is\nplenty.\n\n</what-the-counts-do-not-see>\n\n<when-the-writer-asks-whether-it-reads-as-ai>\n\nMore Writers ask this now, often about their own unassisted prose, and usually\nafter a reader or a submissions guideline spooked them. Answer it from the same\ncounts, and do not guess at provenance — you cannot tell who or what wrote a\npassage, and saying you can is the one answer that does real harm. What you\ncan do is name the habits that make readers say it.\n\nThree are worth looking for, and two of them are already measured.\n\n- **Flat rhythm.** A low `standardDeviation` with the `lengthBuckets` crowded\n  into one or two bands is the strongest of the three. This is the same finding\n  as monotony above; it is what a reader is reacting to when they say prose\n  feels generated. Prose that reads as human usually swings widely — a\n  four-word sentence against a forty-word one — and a coefficient of variation\n  above about 0.5 is comfortable.\n- **Reused wording.** Check `echoes` and run `tabularia_search` on any sentence\n  that felt familiar. A clause repeated near-verbatim in two chapters is the\n  tell a reader meets fastest, and it happens honestly: a Writer drafts the\n  same explanation twice months apart.\n- **Comma-strung inventories.** Four or more parallel items piled into one\n  sentence — \"the maps, the calendars, the relationships, the languages\" —\n  reads as filler wherever it comes from. This is **not** measured, so read for\n  it and never put a number on how many you found.\n\nTwo things a Writer will bring up that are not tells. Em dashes are ordinary\npunctuation in published fiction; say so rather than joining in. And a\ntricolon — three items, balanced — is one of the oldest figures in English\nprose, so do not flag a Writer's rule of three as machine-written.\n\nIf the counts say the passage is fine, say that plainly. \"Nothing here reads\nthat way to me, and here is the spread\" is a complete answer.\n\n</when-the-writer-asks-whether-it-reads-as-ai>\n\n<what-to-say>\n\nThree observations. Each names the count, says what it means here, and points at\na paragraph the Writer can look at.\n\nNot: \"Your standard deviation is 3.2, which is low.\" That is the tool talking.\n\nSomething like: \"The scene runs at sixteen to nineteen words a sentence almost\nwithout a break, which is why the fight reads flat — blocks 210 to 214 are five\nsentences of near-identical length. Two of your other chapters swing between\nfour and thirty.\"\n\nSay what is working too, briefly. A read that is only complaints gets skimmed.\n\n</what-to-say>\n\n<what-not-to-do>\n\n- **Do not propose edits.** This is a read. If they ask you to fix something,\n  that is `line-edit`, and it starts by reading the passage again.\n- **Do not turn this into a checklist.** Every item you raise costs the Writer a\n  decision, and a list of twenty means they act on none.\n- **Do not repeat advice about adverbs, passive voice or \"said\".** They have\n  heard it, it is mostly wrong, and Tabularia's counts are there to replace it\n  with something about their actual book.\n\n</what-not-to-do>\n"
}