Changelog

1.16.0 - 2026-10-20

• Add an opt-in controlled-English register under technical documentation, based on ASD-STE100 Issue 9 (one instruction per sentence, 20/25-word sentence limits, six sentences per paragraph, active voice, simple tenses, three-word noun clusters). Apply it most of the way by default and fully only on request. Never claim conformance without the controlled dictionary, and keep it out of persuasive and voice-matched work.

1.15.2 - 2026-10-20

• Trim the auto-loaded skill description from 121 to 60 words without dropping the implicit-intent triggers covered by tests/trigger-cases.md.

1.15.1 - 2026-10-11

• Codex parity: compatibility, README title and trigger cases now name Claude Code and Codex; "Claude UI Router" becomes UI Router.
• references/model-routing.md defers model, effort and delegation to Work Router and adopts the host-native rule: the parent in the user's current host writes. The dated family matrix (a Sonnet default, Terra explorers, Fable) is removed; the prose-specific effort guidance, one-writer/one-reviewer pattern and reviewer contract stay.

1.15.0 - 2026-09-22

Speaker notes are their own shape of spoken prose.

A full pass on a fifteen-slide customer-council deck followed the spoken register correctly and still produced notes the presenter would not say: the strongest line sat at the end of each slide, segues were captions ("Before the product, thirty seconds on where it sits"), fields were read in order ("the primary artifact is"), and openers reached for aphorisms ("Case is for the work BPMN cannot draw"). Each was a rule the register did not state.

• spoken-register.md gains Speaker notes in a slide deck: lead with the strongest line as the one bold paragraph; the next slide carries the segue, as a full sentence; anchor to the slide's words without reading its fields in order; plain statements, no aphorisms or superlatives about the deck; short beats with a blank line between and bullets for lists; a read-from slide gets intro, one cue per block, outro; no headers, timings or slide numbers in the notes; customer-facing words only.
• SKILL.md's resource table names slide speaker notes as a spoken piece and points at the new section.
• Forward case 42: buried lead plus caption segue plus field read-out in one set of notes.
• Judgment calls stay judgment: the buried lead, the caption-segue and the aphorism are not linted; the section gives the manual test (read every slide's first paragraph in sequence as headlines).

1.14.1 - 2026-09-03

The questions an editor cannot see it is guessing at.

Two forward runs of the same executive update, both following the skill correctly, disagreed on one line: the first cut "a shift in how the whole organisation thinks about data" as unsupported, the second kept it as the author's stance. The draft does not say whether it is a report or an argument, so neither editor noticed it was choosing, and neither asked.

• SKILL.md's pre-pass now names the four questions whose answers are invisible in a written draft, each with its default: purpose (report, persuade, decide); how much of the author's judgment stays; whether a writing sample exists; what is protected. Reader, register and length stay inferred.
• When the pass is too small to ask, the What changed note opens with one line of the assumptions it rested on, so the author corrects before reading the edit rather than after. Eval check 19 now requires it.
• An answered pre-pass question or voice probe is recorded as a standing preference wherever the harness keeps them, so it is asked once.
• Workflow example added showing the two correct edits and why the first line of the note is the repair. Forward cases 40 and 41.
• Forward test on the same executive update: the note opened "Edited as a status report to a VP, your judgments kept where the draft supports them, no writing sample," named the framing line as the one cut resting on that assumption, and said what happens to it under the other purpose.

1.14.0 - 2026-09-03

Every row has an example, in the same file as its table, and tests keep it so.

• examples.md is gone. Each family table in pattern-catalog.md is followed by a worked before/after for every row in it, headed by the row's exact name; the spoken-only rows have theirs in spoken-register.md. 42 examples written for rows that had none (73 in all), including the six creative-prose rows and every assistant-residue row. The three workflow examples (diagnose mode, options when the call is taste, exit check 27) sit in a Workflow examples section at the end of the catalog.
• New row Stock vocabulary. The linter had checked for "delve" and "tapestry" since 1.0 and no row said why; readers rank the words third among the tells they complain about.
• Linter: ROW_FOR maps every check to the catalog row it serves. Three slugs renamed to match their rows: superficial-interpretation, vague-authority, stock-vocabulary. Sixteen rows gain the Linted: marker they were missing.
• New tests/test_catalog_structure.py: every row has an example headed by its name; every example heading names a real row; every lint check maps to a row that exists and every check is mapped; linted rows carry the marker; forward cases that name a row name a real one. The naming drift found by hand in 1.13.1 now fails a test.
• The Unsupported certainty example had split a short sentence and dropped "probably" in the name of preserving hedges. Its better edit is now the sentence unchanged, with the note that tightening it is an option to offer.
• SKILL.md: when no writing sample exists and a small voice choice will recur (a short sentence kept whole or split, contractions, dash style), show the author one sentence both ways and ask. One probe settles the pass.
• Every After block was linted, and a fresh Opus reviewer read all 72 against the eval and every other row. It found 24 problems in 21 examples: eleven invented details (a vendor, a customer, a log, a hand-print), seven dropped hedges or facts, four examples that installed another row's pattern, two Whys that misdescribed the edit. All fixed before release. An example that breaks the rules it sits beside teaches the break.
• Forward test, nine passages, fresh agent holding only this skill: nine pass with every source fact intact. The talk track refused to gloss a screen label from its own words and asked instead, which is the 1.13.0 lesson holding. One weak spot, two quotable closers kept in one memo paragraph.

1.13.1 - 2026-09-03

Naming. The catalog names patterns; examples.md headed each example by its editorial move and named the pattern only sometimes, and not always with the catalog's word. Every example heading is now the exact catalog row name (or the SKILL.md or eval section it demonstrates), every bold reference in an example resolves to a row, and the forward cases that target a row use its name too. "Abstract praise" in the diagnose example becomes Sales varnish, which is the row it was describing.

1.13.0 - 2026-09-03

A structure release. Twelve versions of pattern accretion, most of them from one corpus of keynote talk tracks, had left the general-purpose layers carrying spoken-register weight and three files with overlapping jobs. Nothing the skill teaches has changed; where it lives has.

Layout

• New references/spoken-register.md. Everything spoken in one place: the register rules (moved from voice-and-register.md), the five patterns that only fail out loud (label read cold, speaker meta, trailer cadence, telegraphic speech, stacked precision, moved out of the general catalog), the pre-pass questions, and the --spoken checks. Each of those patterns is harmless in a caption or a spec sheet, so an email editor no longer reads them.
• references/claude-patterns.md is gone. Seven of its patterns were already rows in the catalog, so the file was a second catalog with a different shape. Its unique content moved into the catalog as two new tables, Assistant residue (the seven existing rows plus balanced-by-default stance, correction-first argument, local recap, uniform helpfulness, abstract personification) and Creative and narrative prose (six rows). The catalog is now the one list of written-prose patterns; SKILL.md carries a reference map saying what each file is for.
• SKILL.md's layered edit step 4 no longer names the last five patterns that were added; it states the method and points at the catalog. The Calibrate section points at spoken-register.md instead of restating it.
• sources.md gains two maintenance rules: a new catalog row ships with a worked example and a forward test, and every rule has one home.

Examples and tests

• examples.md goes from 9 worked examples to 22 in the structure pass, 31 after the source check. Every rule family shipped since 1.1 now has a before/after with the reason, plus a spoken example, a set-level example, an options-before-pass example, and a draft that is returned unchanged.
• tests/cases.md goes from 22 forward cases to 39, covering eval checks 23 to 27, --set, the markup round-trip, and the two most recent lint patterns. Two trigger cases added for spoken and set-level work.

Eval

• Check 27: a draft that passes is returned unchanged and said so. The minimum- edit rule had no exit; a 26-item checklist will always find something to do.

Linter

• interface-as-narrator no longer matches the bare pronoun "it" or the verb "says". "The report says the queue is empty" and "she told me it decides" are ordinary English; both fired.
• The three frequency rules (contrastive-definition, deferred-point, mechanism-speak) report once per document with a count and the line numbers, and only from three occurrences. They had printed one line per match, which buried the real findings; the skill's own SKILL.md produced eleven.
• Sentence-level checks (sentence-shape-run, flat-declarative-run, stacked-precision, the spoken paragraph opener) skip headings, list items, table rows, fenced code and frontmatter. A numbered list is not a run of flat declaratives. The per-line scan skips fenced code and frontmatter.
• Twenty-one new tests, 101 pass.

From the sources, checked 2026-09-03

• New row Candor announcement: "Honestly?", "the honest answer is," "here's where I'd push back." The reader-side complaint about 2026 Claude prose that the catalog had no row for; Announced virtue rates the subject, this rates the speaker. Linted on high-precision phrases only; five tests.
• New row Vague connection (Wikipedia, 2026-08-19): "associated with," "linked to" in place of the actual role. Judgment only.
• Nominated significance widened to interpretive metadiscourse ("as you can see," "this distinction matters," "the key insight is"), after no-ai-slop shipped the broader form. Three tests.
• Dash dependency now covers parenthetical asides, which one audit ranked the stronger tell. Em dash default notes that by mid-2026 it had become a Claude-leaning tell rather than a generic one. "full stop" joins the stock vocabulary list.
• The Assistant residue table carries a note on lexical tells and why the fix is never the word.
• Six more examples, for the six most complained-about patterns that had none: not-X-but-Y, the triplet, hedge stacking, bold-term lists, candor announcement, vague connection. 31 examples in all. Six forward cases, 39.
• From hardikpandya/stop-slop (2026-01-13), three rows the catalog lacked: Intensifier padding (twin of Qualification pileup; judgment only, no adverb ban), Universal quantifier ("everyone," "never" where the source says "most"), and Quotable closer (every paragraph ends on a pull-quote; sibling of Recap ending). Four existing rows widen their signal: Nominated significance takes the emphasis crutches ("let that sink in," "make no mistake," both linted) and the pseudo-cleft opener; Question-answer pivot takes "What if I told you"; Negative tail becomes Negative tail and runway; Importance inflation takes the bare declaration of weight ("the stakes are high"). Three of stop-slop's rules were read and declined, and sources.md says which and why.
• sources.md gains a field-reports section, a last-checked date per repo, and the stop-slop entry.

Forward test before release

Seven sloppy passages (executive update, product copy, memo, bio, talk track, a clean control, an assistant-shaped analysis) were edited by a fresh agent holding only this skill. All seven passed the eval with every source fact intact; the control came back unchanged under check 27. One weak line: a screen label glossed from its own words when the source gave nothing else. Label read cold now says to ask in that case.

Hygiene

• The Telegraphic speech entry had a run-on left by a bulk edit. Lint docstring said two whole-piece checks and listed four. This changelog's header was below eight undated entries. README now shows the layout and the three ways to run the linter, and its quick-test prompt no longer keys on "humanize" and "AI slop", which the skill says not to trigger on.

1.12.0 - 2026-09-01

New check: interface-acts-on-itself. "Pick a few processes and the map seeds itself." "The map draws itself." An interface doing something to itself, with no actor named.

The catalog has described this since Interface as narrator shipped, and its own example is "the app keeps score on itself" - but there was no check, so the pattern was only catchable by a human reading the catalog closely. It got past a writing pass and into a published keynote board.

• Catches the reflexive slice, which is the part a regex can hold. The general case stays judgment: the noun and verb lists are open-ended.
• "itself" emphasising a different noun is excluded by refusing a determiner between the verb and the pronoun, so "the frame shows the record itself" stays clean.
• "process" is deliberately not in the noun list. A process genuinely runs, and "the process improved itself" is a claim authors make on purpose.

Measured before shipping: fires on all four controls, zero hits across a 694-field corpus of already-edited board copy. Five new tests, 80 pass.

1.11.0 - 2026-09-01

New pattern: Nominated significance. The line names what matters instead of delivering it. "The owner column is the point." "The re-render is the point." "What matters is how faithful it is."

It survives edits that catch Announced virtue because it does not sound boastful, it sounds structural, and it is doing the reader's noticing for them. The repair is to state the finding: "the owner column is the point" becomes "work that runs every day has no owner." If the finding cannot be stated, the line has nothing to nominate and the problem is the content, not the phrasing.

Shipped as a lint pattern rather than judgment, because it measured clean: 0 hits across a 205-field corpus of already-edited keynote copy, and it catches the flagged control. Four new tests, 75 pass.

1.10.2 - 2026-09-01

A markup round-trip rule, earned by shipping a literal _is_ into a keynote board.

Handing prose to another agent means converting inline markup to a plain-text carrier (<b> to **, <i> to _) and converting it back on the way in. The write-back verified bold anchors and figures, did not verify italics, and put the underscores on screen.

• Eval check 3 now requires an inventory of every inline markup token before and after a round trip - bold, italics, code, links, entities - with the counts compared. Checking only the token you were thinking about is the failure mode.
• Output discipline says the same thing at the point of writing back.

1.10.1 - 2026-09-01

1.10.0 said the compression pattern was unlintable. That was true of the pattern as a whole and wrong about its parts. Four candidate heuristics were measured against the same 40-piece talk-track corpus; three fired only on true positives and ship here behind --spoken, and the fourth was rejected for firing three times, all false.

• --spoken adds compressed-mechanism (a noun-plus-temporal-clause series: "a clock, a warning when it is at risk and an escalation when it breaches"), stacked-object-pronouns ("hand it that"), and paragraph-opens-on-pronoun. Opt-in, because all three are wrong for written prose.
• Rejected and recorded: a verbless comma inventory, which is also the shape of a legitimate pointing beat.
• The catalog entry now says partly linted rather than unlinted, and keeps the reason the general call cannot be automated: no regex can know that "a stage that is meant to take four" is an SLA said the friendly way.

Seven new tests, 71 pass.

1.10.0 - 2026-09-01

Process, not patterns. 1.9.2 named the over-compression defect but left the skill with no way to catch it, because a detector was prototyped and rejected: keyed on a noun-phrase series with a subordinate clause it produced 19 candidates across 40 talk tracks, nearly all deliberate rhetoric, and missed the canonical case, which carries only one comma. What prevents this going forward is a step in the workflow and a gate in the eval, not a regex.

• Say it out loud is now step 8 of the layered edit, for spoken pieces. Deliver every sentence at presentation speed; any sentence you slow down for, re-run, or breathe inside is compressed. Enforced as eval check 23.
• Ask before a substantial pass. Up to four questions whose answers change every line, asked together, each with the default you would otherwise assume. For spoken work: who is speaking to whom, what is on screen behind them, which figures the audience can already read, how long the beat is, and whether any line is dictated or already agreed elsewhere. One earns its place above the rest because it is invisible in the draft: how precise should the numbers be out loud.
• Offer options when the call is taste. Three to five differing in approach rather than phrasing, each naming what it trades away, one recommended, all factually identical, offered before a large pass rather than after. And the lesson that produced this section: when the author picks an option "stylistically" they have chosen the approach and not the words, so fix the flaw they named in it rather than handing the same text back. Eval checks 25 and 26.
• Unlintable patterns are declared, with evidence. When a check cannot be made reliable, say so and hand over the manual test. A noisy check is worse than none: it trains the next reader to skip the output. Any new check gets measured against a real corpus and its false-positive rate reported before it ships.

Eval goes from 22 checks to 26.

1.9.2 - 2026-09-01

A correction to the direction of travel. 1.9.0 and 1.9.1 both push toward tighter sentences, and applied to a talk track that produces spec lines nobody can say out loud.

• New catalog entry Telegraphic speech: a spoken line that compresses a mechanism into a noun-phrase list. "Each main stage carries a clock, a warning when it is at risk and an escalation when it breaches" reads fine and cannot be delivered. The repair is longer on purpose: an opener that buys the room a beat ("Now,"), the subject named rather than pronouned, and the mechanism walked through - "when that SLA is at risk of being breached, Maestro Case will escalate it to the right person automatically."
• The spoken register gains the same rule, plus a guard on the sentence-variety rule so it cannot be read as "shorten everything": cutting every sentence to its shortest form produces the flattest run of the lot, and in a spoken piece the long flowing sentence is usually the one carrying the listener.
• Say the product's own word rather than a friendly paraphrase. If the thing is an SLA, say SLA. Glossing is for column headings and jargon the room cannot parse, not for terms the audience came to hear.

Source: the board owner, on a line this skill's own advice had over-compressed.

1.9.1 - 2026-09-01

Tuning, measured against the corpus that produced 1.9.0. Swept across 40 keynote talk tracks, 1.9.0's new checks fired 12 times on 11 tracks and a Fable review judged 10 of the 11 to be false positives. A check that misfires that often gets ignored, so the checks were narrowed against those cases rather than the rules relaxed.

• stacked-precision no longer counts three things that are not figures: pronominal "one" ("each one", "not one of them", "the one at the top"), alphanumeric identifiers (SR-440, WR-2026-0417, IMP-0005), and elapsed-time markers ("six months on", "three months later").
• flat-declarative-run now knows that a fragment or an imperative is a pointing beat rather than a flat declarative, so short-short-short-long tempo stops reading as a recited list. A verbless label is detected by shape (a comma and no auxiliary or copula anywhere) rather than by verb lookup, after a curated verb list produced a false negative on "Somebody approves the containment." TURN_MARKERS also gained "now", "no longer" and "used to", which carry a then/now contrast.
• The catalog's Stacked precision entry now names its remaining known false positive: a chain of figures that derive from one another and are all on the frame behind the speaker. The check cannot tell a derived chain from an unrelated stack, and pretending otherwise would be worse than saying so.

Result on the same corpus: 12 findings down to 6, all four flat-declarative false positives cleared, and the one true positive still fires. Eight new tests, 64 pass.

1.9.0 - 2026-09-01

Three checks earned by a keynote talk track whose author could hear the problem but could not name it: "you're still have a backwards facing kind of syntax... it's hard to describe" and "I hate the flat declaratives."

• Backwards-facing clause (catalog + spoken register + lint). A clause finished by an earlier one instead of standing on its own: "built for exactly that," "runs the way the map said it could," "where nothing made them wait." Pointer words carry the structure, so the listener has to hold the previous clause to parse this one. Costly in prose, unaffordable in speech, where nobody can look back.
• Flat-declarative run (lint: flat-declarative-run). The monotony check the old sentence-shape-run missed. Three or more consecutive sentences of near-identical length that none of them turns, where each sentence is individually good and the run reads as a list being recited. A sentence carrying a turn - a contrast, a colon, a subordinate move - breaks the run, because that is the prescribed fix.
• Stacked precision (lint: stacked-precision). Three or more consecutive sentences each landing an exact figure. Rounding aloud is the repair, so audibly rounded figures ("call it forty", "close to a hundred") do not count against it. One hard rule added to the spoken register: never say an approximation that contradicts an exact number the audience can read on the screen behind you.

Nine new unit tests, 56 in total.

1.8.0 - 2026-08-28

• scripts/lint_natural_writing.py --set <file>: measure a set of blank-line separated pieces against each other, not each piece against the rules. Flags any opening word, closing pair, or connective shared by more than a fifth of the set. The catalog's Editing a set already demanded this measurement and shipped no way to do it, which is exactly how a careful headline pass left 16 of 40 titles and 34 of 125 key points opening on the same article - the substitution failure the catalog names, invisible from inside any single piece.
• Land callbacks forwards (voice-and-register.md, spoken register): the register already banned act numbers and stage directions, but not the rhetorical callback - "Remember...", "Keep that in mind", "what the last act comes back to" - which looks harmless one at a time and reads as a tic in aggregate. Named as a set-level check.
• Editing a set and SKILL.md now point at --set, so the instruction and the tool that satisfies it are in the same place.
• Four unit tests for the new mode (47 total). A length-spread heuristic was written and then cut: it false-positived on a set that had just been varied on purpose, and the threshold was invented rather than derived.

1.7.0 - 2026-08-28

• Add a Scripts meant to be spoken register (voice-and-register.md): gloss screen labels rather than reading them cold, speak in the room rather than about the talk, keep the screen's own word for each thing, treat bold anchors as navigation the speaker needs (the one register where added bold is earned), and break lines at beat changes. Distilled from a day of live talk-track editing with the author reading every pass.
• Four catalog rows with linter checks where a regex is honest: circular assertion ("the rules ... are these rules"), furniture inventory ("three cards:"), label read cold (catalog-only; register-dependent), speaker meta ("easy to mix up", "hold that pattern", spoken act references).
• Extend synonym cycling: when a screen or document is in view, use its word - "reasons" on screen means not switching to "causes" in the walk.

1.6.0 - 2026-08-28

• Widen interface as narrator to derived artifacts and agency verbs: "the recommendation splits the claim" is the same disease as "the app keeps score" - an artifact doing an actor's verb. The catalog now says the noun list is open-ended on purpose; the linter adds recommendation/finding/proposal/ rationale/brief and splits/decides. Caught in the wild by the reader after the original rule's noun list let it through - a class-boundary miss, not a revert.
• Add trailer cadence: compressed ad-copy rhythm in place of a spoken sentence ("One glance and Sarah has the case"). Let a person do it at normal speed.

1.5.0 - 2026-08-28

• Add interface as narrator: a screen, app, map, record or row performing a human verb, usually with a vague pronoun carrying the sentence - "the app keeps score on itself", "the map admits what it does not know", "each one says who owns it". Say what is on screen, or name the person reading it. The sibling of org-chart actor, and caught the same way: in the wild, three times in one sitting, by the person being read to. show, list and mark are deliberately not flagged - those are what screens do; the tell is attitude.
• Add announced virtue: prose that rates its subject instead of showing it - "and it is honest", "the useful thing about it is", "that matters more than it sounds", "because a picture is not a receipt". Cut the rating and let the evidence carry the judgment. If the evidence is missing, the rating was the only claim being made.

1.4.0 - 2026-08-28

• Add em dash default: use a typed dash (a hyphen, or a hyphen with a space on each side) as the reflexive connector instead of an em dash; keep the em dash for the rare case it is clearly the better mark. New linter check scan_em_dash fires on density across a piece, never on a single instance, matching the existing false-positive guardrail.
• Add clause-shape monotony: break up a run of four or more consecutive sentences in one paragraph that share the same shape, compound/hedged or short and simple. New linter check scan_sentence_shape finds the run and reports its length. Both checks run outside scan(), at the whole-piece or whole-paragraph level rather than per line, since the failure is a pattern across sentences, not a single line.

1.3.0

• Add org-chart actor: a department performing a human verb ("product quality gets the failure record"). The reader pictures nobody; name the people. A function can own or approve; it cannot hear, learn, or remember.
• Add insider jargon: systems vocabulary in reader-facing prose -- "read-only evidence," "system of record," "human-in-the-loop." Say what happens in the reader's world instead. Both caught in the wild on the same paragraph, which is how they earned their rules.

1.2.0

• Add the spec-sheet coda: a finished sentence followed by a verbless list of qualities ("Plain language, no jargon, ready to use as it stands."). It reads as product-blurb summary rather than a claim, and it survived the earlier passes because each fragment is short and none of the existing rules look at what follows a sentence-ending period. Catalogue row, linter rule, three tests.

1.1.0 - 2026-08-25

• Added four pattern-catalog entries: deferred point, contrastive definition, mechanism-speak, and unglossed shorthand.
• Added an "Editing a set" section. Editing many pieces to one standard installs that standard as the next pattern, and each fix tends to substitute for what it removed. Measure the fix, not only the fault, and count words rather than sentences when length matters.
• Added three frequency rules to the linter for the same patterns, documented as corpus-level findings rather than per-line faults.
• SKILL.md now points at "Editing a set" whenever the job covers a set rather than a single piece.

1.0.1 - 2026-08-19

• Broadened auto-invocation from example phrases to semantic writing intent.
• Added paraphrased and implicit trigger cases plus non-writing boundary cases.

1.0.0 - 2026-08-19

• Added Claude Code-native Natural Writing skill and icon.
• Added explicit auto-invocation triggers for editing, drafting, voice matching, anti-slop work, interface copy, and Claude-specific writing residue.
• Added source synthesis, pattern catalog, voice guidance, model routing, examples, deterministic lint checks, and 22 forward-test cases.