@domandigital/craft 0.6.0 → 0.8.0

package/CHANGELOG.md

@@ -1,5 +1,41 @@
 # @domandigital/craft
 
+## 0.8.0
+
+### Minor Changes
+
+- 2b6f823: The house copy standard moves into craft. `COPY.md` ships in the package, `craft copy --gate` applies the house policy (the blocking tier fails the run), and `craft tells list --json` publishes each copy tell's house tier so `copy-check` reads the policy instead of keeping its own list. Catalogue 2026.09.4 adds the density tier (`phrase-density`, `aphorism-density`, `contraction-scarcity`, `sentence-rhythm`, `repeated-sentence`, `heading-shape`, `heading-echo`) and four review tells from research (`ing-tail`, `vague-attribution`, `closing-summary`, `false-range`), all warn. The emoji tell now catches sparkles, the green tick box and other emoji in the older symbol blocks.
+
+  Words inside a URL or a Markdown link target are no longer read as copy, so a slug such as `/seamless-booking` no longer blocks.
+
+- 7765dda: Catalogue 2026.09.6, tuned against the Doman Digital site's live Sanity content. `chatbot-residue` reads a run of ChatGPT citation tokens as one finding, through the invisible private-use characters ChatGPT wraps them in. `placeholder` skips a placeholder inside quotation marks, which is a template being taught. `vague-attribution` skips a claim that names its source or carries a footnote or link. `inline-label-list` fires only when the text after the label is six words or fewer, so a definition list passes.
+- ef5b85b: Catalogue 2026.09.5. Four more review tells: `chatbot-residue` (pasted chat replies, chat citation tokens, links tracked `utm_source=chatgpt.com`), `placeholder` (`[Insert ...]`, lorem ipsum, `TODO_PLACEHOLDER`, including a placeholder that is a whole element), `question-reveal` and `inline-label-list`. New command `craft copy compare <before> <after>`: the preservation gate, listing the protected facts (prices, numbers, dates, times, phones, emails, links, postcodes, names) a rewrite lost or added, exit 1 if any. Exported as `compareFacts` and `protectedFacts`.
+
+## 0.7.0
+
+### Minor Changes
+
+- dbdef56: Signal 2, the counterfactual. `craft null build --brief "..." --out <dir>` asks `claude -p` for about 20 home pages from the brief alone, renders and fingerprints each, and writes `null.json`. `craft audit <url> --null <dir>` scores how typical a page is against them: the share of the model's own pages that sit further out than this one, typical at 0.10 or above, with the choices it shares with most of them. `craft tells harvest <dir>` lists what recurs across null pages and which tell already catches it, so the next generation of tells comes from what the model builds now. `snapshotUrls` in `@domandigital/craft/audit` snapshots several pages through one browser. The rendered reveal check now also reads the text blocks inside a section, so reveals on cards two levels down are no longer read as a still page.
+- f765c69: Signal 3, the estate. `craft estate add <url | snapshot.json> --id <id>` fingerprints a shipped site into `estate.json`; `craft estate compare` lists every pair, or one site against the rest, closest first, and marks siblings: two sites closer together than two pages Claude builds for one brief (0.31, or measured again from `--null`). What a pair shares is named in plain words ("Fraunces headline", "no accent"). `--strict` exits 1 on a sibling. `craft direction propose --estate estate.json` reads the register. The fingerprint no longer takes a small painted patch, such as a floating chat button, for the accent: a painted colour needs a panel's worth of area.
+- f3c3dc0: The character report. `craft report <url | snapshot.json>` reads a site against all three signals and the reason rule (`--repo`, `--null`, `--estate`, `--direction`) and gives a verdict: default, mixed, decided, or unproven when something was not measured. The actions come in retrofit order, one per choice, with every reason that points at it. `craft retrofit` writes them as a Markdown checklist that names the art-direction choice settling each change, shows the decided value, and ends with what to leave alone. `characterReport` and `retrofitPlan` are exported. System font keywords such as `ui-monospace` no longer read as CSS variable names.
+
+### Patch Changes
+
+- 235b558: Export the copy catalogue's word and phrase lists as portable JSON
+  (`wordListsJson()`, checked in at `word-lists.json`), so a non-TypeScript
+  consumer can source from them instead of duplicating them by hand.
+
+  `claude-kit`'s `copy_check.py` hardcoded its own copies of these same lists
+  and had drifted from this package (documented in claude-kit's
+  `house-style/copy-rules.md`, "Note the enforcement gap"). `word-lists.json` is
+  the file that ends the drift: `install/sync-copy-wordlists.sh` on that side
+  copies it in verbatim, the same shape `sync-craft-standard.sh` already uses
+  for `STANDARD.md`.
+
+  No tell's detection or severity changed. Regenerate the checked-in file with
+  `pnpm --filter @domandigital/craft run docs` after editing any list in
+  `tells/copy.ts`.
+
 ## 0.6.0
 
 ### Minor Changes

package/CHARACTER.md

@@ -29,8 +29,8 @@ model defaults move.
 | Signal | Question | Status |
 | --- | --- | --- |
 | 1. Known tells | Does it use a choice on the catalogue below? | **Shipped**: source, copy, and on the rendered page for 18 tells (`craft audit`). |
-| 2. Counterfactual typicality | Would Claude have built this anyway, for this brief? | Planned. About 20 `claude -p` runs per brief; distance from those answers. |
-| 3. Distance from the estate | Does it look like the agency's other sites? | Fingerprint and distance shipped; the estate register is planned. The agency's biggest tell is its clients looking like siblings. |
+| 2. Counterfactual typicality | Would Claude have built this anyway, for this brief? | **Shipped**: `craft null build` generates about 20 pages from the brief; `craft audit --null` scores a page against them. |
+| 3. Distance from the estate | Does it look like the agency's other sites? | **Shipped**: `craft estate add` fingerprints each shipped site into `estate.json`; `craft estate compare` marks siblings. The agency's biggest tell is its clients looking like each other. |
 
 **The reason rule.** Every expressive choice records a `because`, with evidence
 from the client's real world: the shopfront, the van, the trade's own look, the
@@ -43,7 +43,7 @@ adds character rather than only removing tells. It lives in `art-direction.json`
 | --- | --- |
 | 1 | The first wave: indigo, blue-to-purple gradients, Inter, glass, three icon cards. |
 | 2 | What models moved to once the first wave was named: cream grounds and italic serifs, eyebrow chips, bento grids, marquees, intro cinematics. |
-| 3 | Whatever the harvester finds next. |
+| 3 | Whatever `craft tells harvest` finds next in the null models, once a person has read it. |
 
 A report counts findings by generation, so it can say *which* wave a site was
 built in, not only that it has tells.
@@ -185,16 +185,110 @@ against different sets.
 Known limit: a computed font family is the family the stylesheet asked for.
 A face that failed to load is still reported under its name.
 
+## The counterfactual
+
+A list of tells goes stale; the model does not. So signal 2 asks the model
+directly: given this brief and nothing else, what would you build?
+
+```bash
+craft null build --brief "RMP Electrical, electricians based in Uxbridge ..." --out null/rmp
+craft audit https://www.rmp-electrical.co.uk --null null/rmp
+```
+
+`null build` runs `claude -p` about 20 times (`--runs`), from an empty folder
+with no tools, settings, skills or MCP servers, so each page is what the model
+builds from the brief alone. The prompt is recorded in `null.json` and says
+nothing about how the page should look. Each page is rendered, snapshotted and
+fingerprinted like any other. A reply that is a note instead of a page is kept
+as `NN.reply.txt` and generated again. The build is resumable: run the same
+command to fill in anything that failed. `--brief` can come from
+`art-direction.json` with `--direction`.
+
+**Typicality** is a page's mean fingerprint distance to its three nearest null
+pages, ranked against the same measure for every null page against the
+others. A score of 0.40 means 40% of the model's own pages sit further from
+the rest than this one does. A page is typical at 0.10 or above, so a page the
+model built is flagged nine times in ten by construction, and the threshold
+comes from the model's own spread rather than a number picked here. The report
+lists what the page shares with most of the null ("display Inter, 14 of 20"),
+which is where to start changing it.
+
+**The harvest** reads null models the other way round. `craft tells harvest
+calibration/null` lists the choices that recur across the null pages (faces,
+accent families, grounds, button shapes, openings, phrases) and says which
+tell already catches each. The ones nothing catches are the candidates for
+generation 3. A phrase has to turn up across more than one brief, so a trade's
+own words do not count, and navigation labels, form labels and the legal footer
+are left out as page grammar. Nothing is added to the catalogue automatically:
+a candidate becomes a tell with a flag case and a pass case like every other.
+
+Known limits: a user-level `CLAUDE.md` on the machine that builds the null
+still reaches the model. The null is only as current as its build date; rebuild
+it when the model changes.
+
+## The estate
+
+The agency's biggest tell is not on any one site. It is its clients looking
+like each other. `estate.json` holds one fingerprint per shipped site:
+
+```bash
+craft estate add https://www.rmp-electrical.co.uk --id rmp --client "RMP Electrical"
+craft estate compare            # every pair, closest first
+craft estate compare rmp        # one site against the rest
+craft estate compare https://staging.example.test   # a new build before it ships
+```
+
+**Siblings** are two sites closer together than two pages Claude builds for
+the same brief usually are: the median of that distance over the null models,
+0.31 on 2026-09-23 (`--null` measures it again from any null models given).
+If a barber and an electrician look more alike than two drafts of one barber,
+neither look was chosen. The report says what the pair shares in plain words
+("Fraunces headline", "no accent", "pill buttons"), which is where to start
+pulling them apart. A sibling is a warning; `--strict` exits 1 on one, for a
+pipeline that has decided to hold the line.
+
+`craft direction propose --estate estate.json` reads the register too, so a
+proposed accent steers away from one a sibling already uses.
+
+## The character report
+
+`craft report` reads one site against everything above and gives a verdict:
+
+```bash
+craft report https://www.rmp-electrical.co.uk --repo . --null null/rmp --estate ../estate.json
+craft retrofit https://www.rmp-electrical.co.uk --repo . --null null/rmp --estate ../estate.json --out RETROFIT.md
+```
+
+| Verdict | When |
+|---|---|
+| default | Two or more signals raised: three or more design tells, typical of the brief's null, a sibling in the estate, or fewer than five choices decided with a reason. |
+| mixed | One raised. |
+| decided | None raised, and all four measured. |
+| unproven | None raised, but something was not measured. Unverified is never a pass. |
+
+The actions come in the order a retrofit works: decide first, then type,
+colour, shape, effects, motion, how sections look, and copy. Where a tell, the
+null model and a sibling all point at one choice, they are one action with
+three reasons. Sharing a white page or a still one with the null is never
+an action: that would reward strangeness.
+
+`craft retrofit` writes the same as a checklist. Each change names the
+art-direction choice that settles it, and shows the decided value and its
+reason, or says it waits on that decision. It ends with what to leave alone:
+navigation, the order of sections, where the call to action sits.
+
 ## The house copy rules live here
 
-Since catalogue `2026.09.2` the rule lists of claude-kit's
-`house-style/copy-rules.md` are carried by tells in this catalogue:
-`ai-phrase`, `plainer-word`, `plain-english`, `buzzword`,
-`negative-reassurance`, `em-dash`, `emoji`, `not-just-but`, `no-x-no-y` and
-`no-x-badge` are its blocking tier; `review-phrase` and `vague-word` its review
-tier. craft ships all of them as `warn`. `copy-check` in claude-kit is a thin
-wrapper over `craft copy --json` and decides which tells fail a commit, so the
-lists exist once, here, with a test per entry.
+The house copy standard is [COPY.md](./COPY.md), in this package. Its
+blocking tier is carried by ten tells here: `ai-phrase`, `plainer-word`,
+`plain-english`, `buzzword`, `negative-reassurance`, `em-dash`, `emoji`,
+`not-just-but`, `no-x-no-y` and `no-x-badge`. craft ships every tell as
+`warn`; the house policy in `src/character/house.ts` says which fail a commit,
+and `craft copy --gate` applies it. `copy-check` in claude-kit is a thin
+wrapper that reads the same policy from `craft tells list --json`, so the
+lists and the policy exist once, here, with a test per entry. Since catalogue
+`2026.09.4` the density tier and four research tells are here too, and since
+`2026.09.5` four more, including `chatbot-residue` and `placeholder`. All review.
 
 The swap was checked against the old checker on 2,570 files across the seven
 client repos (2026-09-23). The old checker found 193 issues, craft finds every
@@ -212,9 +306,10 @@ reason that rule exists.
 The targets, recorded here with the date once each has been run:
 
 - **AI set** (pages generated by `claude -p` from about 20 briefs): 90% or more
-  flagged as typical or tell-heavy. *Not yet run.*
+  flagged as typical or tell-heavy. **First run 2026-09-23: 19 of 20 (95%)**, below.
 - **Human set** (the ten references in claude-kit's `reference-set.md`): zero
-  block hits and low typicality. *Not yet run.*
+  block hits and low typicality. **First run 2026-09-23: zero block hits, none
+  typical**, below.
 - **Estate baseline** (all seven client repos and live URLs): every hit read by
   a person before any rule blocks. **First run 2026-09-23**, below.
 
@@ -240,7 +335,7 @@ blocks; it is for a person to read.
 | `glass-panel` | 9 / · | 45 / · | 13 / · |  | 1 / · |  |  |
 | `hero-then-proof` | 9 / · | 1 / R | · / R |  | 1 / · | 1 / · | 2 / R |
 | `icon-tile-grid` | 1 / R | 9 / · | · / R |  | · / R |  | 1 / R |
-| `reveal-everywhere` | 1 / · | 1 / · |  |  |  |  |  |
+| `reveal-everywhere` | 1 / · | 1 / · |  |  |  | · / R |  |
 | `pill-everything` | 1 / R | 1 / · | 1 / · |  |  |  | 1 / R |
 | `shadcn-dump` |  | 1 / · |  |  |  |  |  |
 | `cream-palette` |  | 3 / · |  |  |  |  | 18 / R |
@@ -258,15 +353,17 @@ What it shows:
 - **The three survey findings reproduce where they can.** Sensphere's violet
   is found in source (18) and on its buttons; MMM's intro, glows and glass are
   in source and its intro and glows on the page; MMM reveals 60% of the
-  sections below its fold on scroll, three of five, one short of the rendered
-  threshold, and its source is over the reveal limit. Rise & Bloom's glass
+  sections below its fold on scroll, all three of them, one short of the
+  four the rendered rule needs, and its source is over the reveal limit. Rise & Bloom's glass
   hero and icon grid could not be checked: that repo is not attached to this
   session.
 - **Every site ships a reflex face.** Twelve of fourteen site-face pairs are
   on one of the two lists. Fraunces sets the headline on DD, HJ Beauty and RMP.
-- **DD and RMP are the closest pair** (fingerprint distance 0.46: Fraunces
-  headlines, pill buttons, a green accent). Chair and Blade is furthest from
-  everything (0.59 to 0.77). This is signal 3's first measurement.
+- **DD and RMP were read as the closest pair** (0.46) on the first run, with
+  a green accent in common. DD's "green" was its floating WhatsApp button,
+  48px square, the only saturated paint on a near-black and white page. The
+  fingerprint now needs a panel's worth of area before a painted colour
+  counts as the accent. The corrected pairs are under *The estate*, below.
 - **DD's violet is its brand.** Its 36 source hits wait for the exception in
   its `art-direction.json`.
 
@@ -280,16 +377,164 @@ False positives found and fixed in this run:
   `cormorantGaramond`) are now read back to the face's name, or every Next.js
   site would have passed the font tells.
 
+- The rendered reveal check read only a section and its direct children.
+  Found through the null models (below): a reveal on the cards two levels
+  down read as a still page. It now reads the text blocks inside a section
+  too, ignoring anything fixed. Re-run the same day: Chair and Blade, which
+  has a `RevealWrapper`, now shows `reveal-everywhere` on the page (six of
+  nine sections below the fold).
+
 Not yet read by a person: the source counts for `icon-tile-stack` (MMM 93)
 and `glass-panel` (MMM 45) look high and are the first to check.
 
+### Signal 2 first read, 2026-09-23
+
+Seven null models, one per estate brief, 20 pages each (140 pages), built
+with `craft null build` from the briefs in `calibration/briefs.json`. Eight
+replies were notes instead of pages and were generated again (kept as
+`*.reply.txt`). Everything is in `calibration/null/`; `score.mjs` reproduces
+the numbers and writes `results.json`.
+
+**AI set.** Twenty briefs the null never saw (a dentist, a bakery, a
+landscaper and so on), one page each, scored against all 140 null pages.
+Flagged means typical (0.10 or above) or tell-heavy (three or more distinct
+design tells), a rule fixed before any result was read.
+
+| | Pages | Share |
+|---|---|---|
+| Flagged | 19 of 20 | 95% (target 90%) |
+| Typical on its own | 13 of 20 | 65% |
+| Tell-heavy on its own | 16 of 20 | 80% |
+
+The one miss is the landscaper: dark forest green on a cream ground,
+Cormorant Garamond over Karla, square buttons. Two tells (`cream-palette`,
+`reflex-font-2`), one short of tell-heavy, and a score of 0.01 against the
+pooled null because no estate brief is a garden business. Neither signal alone reaches the
+target; together they do, which is the reason for having more than one.
+
+**The model's look depends on the brief.** Each brief's null pages scored
+against the other six briefs' pages are typical only 2 to 13 times in 20.
+A barber and an electrician get different pages. So a site is scored against
+the null built from its own brief, and pooling briefs is a weaker test.
+
+**The estate against its own brief's null.** Score, then what it shares with
+most of the null.
+
+| Site | Score | Typical | Shares with the null |
+|---|---|---|---|
+| DD | 0.15 | yes | pill buttons (19 of 20), Fraunces headline (13 of 20) |
+| Harrison James | 0.10 | yes, at the line | Inter body (18 of 20), yellow accent (16 of 20), white ground |
+| MMM | 0.05 | no | white ground, radial glow (13 of 20), Cormorant Garamond (9 of 20) |
+| HJ Beauty | 0.05 | no | white ground, Fraunces (5 of 20), marquee (5 of 20) |
+| Chair and Blade | 0.05 | no | square buttons (18 of 20), Bebas Neue (9 of 20), dark ground |
+| RMP | 0.05 | no | eyebrow chip (14 of 20), Manrope body |
+| Sensphere | 0.00 | no | radial glow (10 of 20), Inter body |
+
+Read with care: "not typical" says the site is not what Claude builds from
+the brief today. It does not say the site was decided. Sensphere scores 0.00
+and still carries five catalogue tells on the page, violet among them: Claude
+no longer builds that look for this brief, but it is still a known tell. That
+is what signal 1 is for. DD and Harrison James are the
+two a model would most nearly have built unprompted.
+
+**The harvest** (`calibration/null/harvest.txt`). No face outside the two
+reflex lists is chosen by a quarter of the null pages, so the font lists still
+cover what the model picks today: Inter 69 of 140, Fraunces 51, Poppins 36.
+The catalogue tells the model still produces most: `reflex-font-2` 103 of
+140, `glass-panel` 100, `reflex-font` 94, `cream-palette` 81,
+`pill-everything` 79. Candidates nothing catches yet, for a person to read:
+
+- A coral or terracotta accent (around `#e77a63`): 66 of 140, five briefs.
+- Phrases across five or more briefs: "done properly" (33 pages, six briefs),
+  "exactly what" (27, all seven), "who actually" (24, five), "within one
+  working day" (32, five), "rather than" (27, six).
+- The opening "hero > hero > cards" (54 of 140) is how the section reader
+  splits a nav band from the hero, not a model habit. Not a candidate.
+
+`glass-panel` at 100 of 140 includes source hits on blurred sticky navs, the
+same false positive fixed on the rendered path. The source rule needs the same
+fix before that count means anything.
+
+### Signal 3 first read, 2026-09-23
+
+`craft estate add` on the seven live snapshots, then `craft estate compare
+--null calibration/null/*` (the line measured again: 0.31). Register and output
+in `calibration/estate/`.
+
+| Pair | Distance | Shares |
+|---|---|---|
+| DD + HJ Beauty | **0.28, siblings** | no accent, Fraunces headline, no reveals |
+| DD + RMP | 0.49 | Fraunces headline, grey ground, pill buttons |
+| Harrison James + Sensphere | 0.50 | white ground, square buttons, no reveals |
+| HJ Beauty + RMP | 0.56 | Fraunces headline |
+| Chair and Blade + anything | 0.60 to 0.86 | at most a marquee (with HJ Beauty) |
+
+One sibling pair in twenty-one, and it is the agency's own site with one of
+its clients: the same Fraunces headline over a near-black and white page with
+no colour of its own. Fraunces sets the headline on three of the seven. Chair
+and Blade shares nothing with anyone but a marquee. Not yet read by a person.
+
+### Human set, 2026-09-23
+
+The references in claude-kit's `reference-set.md`, live home pages, scored as
+the AI set is: typicality against all seven null models pooled, and the
+design tells on the rendered page. The tenth reference is "any
+generated-looking site" and is left out by definition. Snapshots and
+`score.mjs` in `calibration/human/`.
+
+| Site | Score | Tells on the page | Flagged |
+|---|---|---|---|
+| Linear | 0.00 | 7: violet, grid, icon grid, marquee, glow, Inter, hairline shadow | tell-heavy |
+| Stripe | 0.00 | 5: violet, blue-to-purple, glass, icon grid, glow | tell-heavy |
+| DD | 0.00 | 4: icon grid, pills, glow, Fraunces | tell-heavy |
+| animations.dev | 0.00 | 2: marquee, Inter | no |
+| Utopia | 0.00 | 2: hero then proof, pills | no |
+| APCA (Myndex) | 0.00 | 0 | no |
+| rauno.me | 0.01 | 0 | no |
+| StarrLuxury | 0.01 | not measured: a preview gate, not the site | |
+| candy-aesthetics | 0.05 | not measured: "This preview has ended" | |
+
+- **Zero block hits and none typical**: both targets met. Every score is 0.05
+  or under.
+- **Three of the seven measured are tell-heavy.** Linear and Stripe are where
+  the first-wave look came from: the indigo, the glow, the gradient were
+  theirs before they were every model's. The tell list describes them too,
+  which is the reason no single signal decides the verdict.
+- **The comparison is not like for like.** The null models are small-business
+  briefs; these are software and portfolio sites, so a low score is easier to
+  reach. A human set of small-business sites nobody generated is the better
+  test and is still to find.
+- Both house demos had expired or gated previews on the day. They are not
+  measured, not clean.
+
+### Character report first read, 2026-09-23
+
+`craft report` on each live snapshot with its repo, its own brief's null and
+the register. Reports and retrofit plans are in `calibration/report/`.
+
+| Site | Verdict | Raised | Actions |
+|---|---|---|---|
+| DD | default | tells (12), typical (0.15), sibling of HJ Beauty | 26 |
+| HJ Beauty | default | tells (6), sibling of DD | 18 |
+| MMM | mixed | tells (18) | 27 |
+| Sensphere | mixed | tells (9) | 18 |
+| Chair and Blade | mixed | tells (7) | 11 |
+| RMP | mixed | tells (8) | 18 |
+| Harrison James | mixed | typical (0.10) | 9 |
+
+No site has an `art-direction.json`, so the reasons are not measured on any of
+them and none can be *decided* yet. The first change after deciding is the
+same on all seven: the display face. It is the estate's common default, and
+the one choice the retrofits (phase G) should settle first. Not yet read by a
+person.
+
 ## The catalogue
 
 Generated from the package. Run `pnpm --filter @domandigital/craft run docs`
 after changing an entry; a test fails until you do.
 
 <!-- craft:catalogue:start -->
-Catalogue version `2026.09.3`, 39 tells.
+Catalogue version `2026.09.6`, 54 tells.
 
 | Id | Gen | Surface | Severity | Tell | Why it is a default |
 | --- | --- | --- | --- | --- | --- |
@@ -332,4 +577,19 @@ Catalogue version `2026.09.3`, 39 tells.
 | `no-x-badge` | 2 | copy | warn | 'No X' badge | 'No catch.', 'NO JARGON GUIDE': a short standalone 'No X' reads as a slapped-on kicker, and worse when the same one is reused across pieces. |
 | `staccato-triplet` | 2 | copy | warn | Staccato triplet | Three fragments in a row ('Fast. Friendly. Local.' or 'No fuss. No jargon. Just results.') is the second wave's favourite rhythm. |
 | `where-x-meets-y` | 2 | copy | warn | 'Where X meets Y' | 'Where luxury meets comfort' is a tagline shape that fits every business and so describes none. |
+| `chatbot-residue` | 1 | copy | warn | Chatbot residue | 'Certainly! Here's a revised version', 'as an AI language model', a citeturn0search0 token, a link tracked utm_source=chatgpt.com: text that exists only because a chat reply was pasted whole. This is evidence, not style. |
+| `placeholder` | 1 | copy | warn | Unfilled placeholder | '[Insert testimonial]', '[Your Name]', lorem ipsum, TODO_PLACEHOLDER: scaffolding a reader must never see, and the clearest sign a page shipped before anyone read it. |
+| `question-reveal` | 2 | copy | warn | Staged reveal | 'The result? Faster growth.', 'What does this mean for you? It means...': the writer asks the reader's question for them and answers it. An infomercial hook, and a favourite of generated copy. |
+| `inline-label-list` | 2 | copy | warn | Bold-label bullets | '**Speed:** Faster pages' three times in a row: a short label restated in a few words, the list shape generated copy defaults to. A definition list, with a real explanation after each label, is not this. |
+| `ing-tail` | 2 | copy | warn | Empty -ing tail | '..., ensuring peace of mind', '..., highlighting our commitment': a participle tacked on the end that claims significance and names no mechanism. One of the most common shapes in the Wikipedia guide to AI writing. |
+| `vague-attribution` | 1 | copy | warn | Unnamed source | 'Studies show', 'experts agree', 'it is widely known': a claim credited to nobody. A reader cannot check it, and the house proof rule forbids a claim nobody can check. |
+| `closing-summary` | 1 | copy | warn | Closing summary | 'Overall,', 'In conclusion,', 'In short,': a paragraph that announces it is summing up, then restates what the reader has just read. |
+| `false-range` | 1 | copy | warn | False range | 'From first-time buyers to seasoned investors alike' names two ends of no real scale, to sound as if it covers everyone. It describes nobody. |
+| `phrase-density` | 2 | copy | warn | Leaned-on phrase | 'Rather than', 'actually', 'That is...' openers, participle triads: each is ordinary English once, and a document that uses one far above its rate has one move and repeats it. |
+| `aphorism-density` | 2 | copy | warn | Aphorism cadence | Two clipped sentences alone on a line ('All of this already exists. The job is making it findable.') land once. Closing every section, they are a cadence a reader learns to hear. |
+| `contraction-scarcity` | 1 | copy | warn | No contractions | A long document with almost no contractions reads stiff and machine-made. People write 'we're' and 'don't'. |
+| `sentence-rhythm` | 2 | copy | warn | Metronomic rhythm | Sentence after sentence of the same length reads generated. A person's sentences run long, then short, then long again. |
+| `repeated-sentence` | 1 | copy | warn | Repeated sentence | The same sentence on two pages, or the same nine words lightly edited, is one of the strongest signs a document was assembled rather than written. |
+| `heading-shape` | 2 | copy | warn | One heading shape | Headings that all share one shape ('X, and Y', 'X, because Y', 'Keep, rewrite, consolidate, retire') read as a template filled in section by section. |
+| `heading-echo` | 2 | copy | warn | Heading echoed | A heading repeated word for word as the first line under it spends the reader's attention twice on the same words. |
 <!-- craft:catalogue:end -->