Give impeccable four measured modes instead of adjectives

The skill described production-grade typography, motion and spacing, which is a
description of a bar rather than the bar. The one head-to-head anyone has run
found prose instruction without retrieval made outcomes worse than no
instruction at all, 9.94% against a 6.08% baseline, while the same instruction
with retrieval context reached 1.82%. Telling a model to be tasteful is
net-negative; handing it a measured type scale is not.

Four modes, measured 2026-08-23 from live sites via computed styles, pixel
histograms, canvas autocorrelation and font-outline run-length measurement.
Each signature is written as implementable CSS rather than named:

pixel      four feTurbulence filters swapped by steps(1) at 400ms/4 = 10fps,
           the traditional animation boil rate, on exactly one element
plate      a display face quantized to 32 units per em, so step = fontsize/32
           and scale alone flips it between Renaissance serif and CRT bitmap
editorial  a sticky 100svh hero overscanned to 1.17x that never moves a pixel,
           with an opaque curtain scrolling over it; no parallax JS at all
scrollfield  animation-timeline with stagger encoded as animation-range offsets
           rather than delays, so mid-transition elements coexist at different
           opacities while scrolling; no motion library on the page

Invariants are separated from divergences because the divergences contradict
each other -- 2px base against 4px, grain against none, binary radius against a
three-step set -- and averaging resolves each contradiction toward the default.

Opt-in via .impeccable/brand.json, which ui-gate/rules/tokens.sh already reads.
Absent, nothing fires.

Claude Design recorded as a publishing lane: /design-sync pushes a design system
up, the return trip is a manual export, there is no in-session pull.

Author: itsvedantkumar

.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
   "plugins": [
     {
       "name": "vstack",
-      "version": "1.22.0",
+      "version": "1.23.0",
       "source": "./claude",
       "description": "28 skills that fire without a slash command, 14 agents, 15 commands, and the session hook that routes situations to skills. Most skills are ported from pstack and Superpowers — see claude/skills/ATTRIBUTION.md for per-skill source and license.",
       "category": "workflow"

CHANGELOG.md

@@ -6,6 +6,40 @@ Versions follow [semver](https://semver.org). The version lives in two manifests
 
 ## Unreleased
 
+## 1.23.0 — 2026-08-23
+
+**Four visual modes, measured rather than described.** `impeccable` described the bar in
+adjectives, which cannot produce it. The only head-to-head anyone has run on config-layer
+interventions found prose instruction without retrieval made outcomes worse than none — 9.94%
+regression rate against a 6.08% baseline — while the same instruction with retrieval context
+reached 1.82%. So the bar now ships as numbers read off live sites: computed styles, pixel
+histograms, canvas autocorrelation, font-outline measurement.
+
+`reference/modes/` carries `pixel`, `plate`, `editorial` and `scrollfield`, each with its
+signature mechanism written out as implementable CSS. What held across all four is separated from
+what distinguishes them, because the distinguishing properties contradict each other directly —
+base unit 2px against 4px, grain against none, binary radius against a three-step set — and
+averaging them resolves every contradiction toward the stock default.
+
+The invariants are enforceable claims, not taste: weight 400 for all display and body with
+hierarchy carried by size and tracking; tracking as a monotonic function of size locked in `em`,
+crossing zero at 15–20px; line-height inverting with size; one accent hue or none; motion in two
+bands with nothing between 300 and 420ms.
+
+**Opt-in per project.** A `.impeccable/brand.json` selects the mode and declares tokens.
+`ui-gate/rules/tokens.sh` already reads its `type.scale` and fails a build off it. With no such
+file nothing fires. A skill that decides on its own whether something deserves to be beautiful
+would fire wrongly and constantly; a file in the repo is a decision somebody made.
+
+**Claude Design documented as a publishing lane.** `/design-sync` and `/design-login` are built
+into the CLI binary at 2.1.239 with a seven-operation `DesignSync` tool. Flow is one way — push a
+design system up, export by hand to come back — so it is not a generation lane and should not be
+planned around as one. Where the account lacks access, say so and tell the user to enable it
+rather than working around the absence.
+
+Two defects were recorded from the references rather than copied: a 12px consent line failing AA
+at 4.34:1, and a scroll-driven site whose `prefers-reduced-motion` coverage misses its own hero.
+
 ## 1.22.0 — 2026-08-23
 
 **Register rule tightened after it failed once.** v1.21.0 banned commentary on the facts and did

README.md

@@ -117,7 +117,7 @@ while measuring nothing. The first five:
 - Two mutations stopped matching after a prose rewrite and reported the checks as unfalsifiable.
   The suite now distinguishes "this check is weak" from "this mutation landed nowhere".
 
-The v1.22.0 audit found thirteen more, and they fall into four shapes worth naming:
+The v1.23.0 audit found thirteen more, and they fall into four shapes worth naming:
 
 - **An anchor in prose that a prose edit moved.** Check 18's comparison of the README's
   context-cost figure was guarded by a grep for a sentence the same commit reworded into a table
@@ -163,9 +163,9 @@ cd vstack
 **Pinned to a release, reading the script first.**
 
 ```bash
-curl -fsSL https://raw.githubusercontent.com/itsvedantkumar/vstack/v1.22.0/bootstrap.sh -o bootstrap.sh
+curl -fsSL https://raw.githubusercontent.com/itsvedantkumar/vstack/v1.23.0/bootstrap.sh -o bootstrap.sh
 less bootstrap.sh                        # about 100 lines
-VSTACK_REF=v1.22.0 bash bootstrap.sh     # installs that tag, not main
+VSTACK_REF=v1.23.0 bash bootstrap.sh     # installs that tag, not main
 ```
 
 Check 24 fails if a version named in these docs is not a tag that exists. It was added after this

claude/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "vstack",
-  "version": "1.22.0",
+  "version": "1.23.0",
   "description": "Skills that fire on the situation instead of a slash command, plus the subagents, commands, and session hook that make them fire. Verification gates, parallel fan-out, code review, and writing discipline.",
   "author": {
     "name": "Vedant Kumar"

claude/skills/impeccable/SKILL.md

@@ -33,6 +33,56 @@ The mode names what the visitor's success looks like on this surface.
 
 Choose the mode from the requested surface, not the product, and persist it only in that surface brief. A tool's landing page is still Persuade; a fashion house's documentation is still Read; a docs index is Read, not Persuade. See [new-work.md](reference/new-work.md) for new surfaces and [operate.md](reference/operate.md) for deeper Operate/Read guidance.
 
+## Visual modes, measured
+
+The four surface modes above answer *what the visitor is doing*. They do not answer *what it looks
+like*, and a description of visual quality cannot produce it: the one head-to-head anyone has run
+found prose instruction without retrieval made outcomes worse than none (9.94% regression rate
+against a 6.08% baseline), while the same instruction with retrieval context reached 1.82%.
+
+So the visual bar is carried by numbers read off live sites rather than adjectives.
+[reference/modes/](reference/modes/) holds four, measured 2026-08-23 from computed styles, pixel
+histograms and canvas autocorrelation:
+
+| mode | reference | signature | base unit | texture |
+|---|---|---|---|---|
+| [pixel](reference/modes/pixel.md) | talk2hug | 4 SVG turbulence filters swapped at 10fps | 4px | grain .52 screen |
+| [plate](reference/modes/plate.md) | perseus | display face quantized to 32 units/em | **2px** | grain .06 overlay |
+| [editorial](reference/modes/editorial.md) | forge | sticky pinned hero, the page scrolls over it | 4px | **none** |
+| [scrollfield](reference/modes/scrollfield.md) | lighthouse | `animation-timeline`, reveals bound to scroll not to a clock | **2px** | **none** |
+
+[reference/modes/README.md](reference/modes/README.md) carries what held across all three —
+weight 400 everywhere, tracking as a monotonic function of size locked in em, line-height
+inverting with size, one accent hue or none, motion in two bands with nothing between 300 and
+420ms. Treat a violation of those as a defect. The divergences are the modes, and they contradict
+each other directly on base unit, texture and radius; averaging them resolves every contradiction
+toward the stock default.
+
+**These apply only where a project opted in.** A `.impeccable/brand.json` in the target repo
+selects the mode and declares the tokens; `ui-gate/rules/tokens.sh` reads its `type.scale` and
+fails a build off it. With no such file, nothing here fires and the surface is built normally.
+Schema and field-by-field rationale:
+[reference/modes/brand.schema.json](reference/modes/brand.schema.json).
+
+## Claude Design
+
+Claude Code ships `/design-sync` and `/design-login` built into the CLI binary (verified at
+2.1.239), backed by a `DesignSync` tool with `list_projects`, `create_project`, `get_project`,
+`list_files`, `write_files`, `delete_files`, `finalize_plan`. It detects Storybook or bare-package
+repo shape, builds and grades component previews, and maintains `.design-sync/` in-repo.
+
+Direction is one way: **Claude Code pushes a design system up to claude.ai/design.** Coming back is
+a manual Export, "Hand off to Claude Code". There is no in-session pull and no public API, so this
+is a publishing lane, not a generation lane. Do not plan around generating a Design artifact from
+here.
+
+Requires the feature enabled on the account. If `/design-login` reports no access, say so and tell
+the user to enable Claude Design for their account — Pro, Max and Team have it; Enterprise has it
+off by default and an admin must turn it on. Do not work around the absence.
+
+Do not copy a signature move into a project that did not ask for that mode. One jittering element
+reads as deliberate; two read as a broken renderer.
+
 ## Commands
 
 | Command | Category | Description | Reference |

claude/skills/impeccable/reference/modes/README.md

@@ -0,0 +1,76 @@
+# Measured modes
+
+Three reference sites, measured rather than described: computed styles read from a live browser,
+pixel histograms, autocorrelation on canvas output. Adjectives were excluded on purpose. The one
+head-to-head anyone has run on config-layer interventions found prose instruction without
+retrieval made outcomes worse than no instruction (9.94% regression rate against a 6.08%
+baseline) while the same instruction with retrieval context cut it to 1.82%. A mode file carries
+numbers for that reason.
+
+Sources: talk2hug.com, perseus.computer, forgeresidency.com, lighthousehq.com, measured 2026-08-23.
+
+## What every one of them does
+
+These held across all three. Treat a violation as a defect, not a choice.
+
+**Weight carries no hierarchy.** `font-weight: 400` for every display and body element. Forge uses
+500/600 on exactly two buttons and the wordmark. Hierarchy comes from size and tracking.
+
+**Tracking is a monotonic function of size, locked in `em`.** Negative on display, zero in the
+middle, positive on small uppercase. Measured curves:
+
+| site | at display | crossover | at eyebrow |
+|---|---|---|---|
+| forge | −0.12em @ 400px | 0 @ ~20px | +0.22em @ 11px |
+| perseus | −0.04em @ 124px | 0 @ 15–18px | +0.16em @ 10px |
+| hug | −0.05em @ 48px | 0 @ 16px | +0.24em @ 14px |
+
+Locked in `em` so the ratio survives a fluid resize untouched. Forge's mobile display is exactly
+0.50× desktop with byte-identical tracking; that is why it does not read loose.
+
+**Line-height inverts with size.** Display 0.88–0.95. Mid 1.35–1.5. Body 1.5–1.75. No exceptions
+in any of the three.
+
+**One accent hue, or none.** Perseus `#647fff`. Forge `#16a85a`. Hug has no type accent at all and
+reserves pure `#ffffff` for a single line of stats. Everything else is a neutral ramp plus alpha.
+
+**Ink is never pure white on a dark ground.** Hug `#f7f4ec` (4.3% warm). Perseus `#ecebf2` (cool).
+Ground is never pure black: `#050505`, `#09090b`, `#060708`.
+
+**Tight groups inside enormous silence.** The ratio is the signature, not any single value:
+forge 7:1 (16px intra, 112–128px section), perseus 16:1 (1:1.5:3:5:16), hug 24:1 (8px to 192px).
+
+**Measure under 60ch.** Forge 42–58ch. Perseus 55–62ch for prose. Hug 32.6ch. Machine output is
+exempt and runs long on purpose (perseus terminal transcript at 134ch).
+
+**Emptiness is the product.** Text ink coverage: hug 1.22% of hero pixels, forge 12–27% per
+section, perseus ~20% of hero. Nothing here is dense.
+
+**Motion has two bands and a hole.** Hover and colour 130–200ms. Entrance 420–760ms. Nothing
+between 300 and 420ms in any of the three. Stagger steps are small and explicit: 45ms (forge),
+75/80ms (perseus).
+
+**prefers-reduced-motion is honoured structurally, not gestured at.** Hug removes the filter
+entirely rather than freezing a distorted frame. Forge defines its marquee only inside
+`@media (prefers-reduced-motion: no-preference)` so it never starts. Copy that, not a blanket
+duration clamp.
+
+## Where they diverge, and what that makes them
+
+| | pixel | plate | editorial | scrollfield |
+|---|---|---|---|---|
+| reference | talk2hug | perseus | forge | lighthouse |
+| base unit | 4px | **2px** | 4px | **2px** (8pt conformance 13%) |
+| grain | feTurbulence `screen` .52 | feTurbulence `overlay` .06 | **none** | **none** |
+| radius | 0 | 8 / 12 / 28 | **0 or 9999 only** | — |
+| elevation | `blur(14px)` | `blur(56px)`, shadow `0 28px 80px/.76` | **zero shadows**, `ring-1` | 31× `backdrop-filter` |
+| display tracking | −0.05em | −0.04em | −0.12em | **`normal`, even at 80px** |
+| motion | CSS filter swap | mount-time entrances | IntersectionObserver | **`animation-timeline`, no library** |
+| signature | `pixel.md` | `plate.md` | `editorial.md` | `scrollfield.md` |
+
+`scrollfield` is the one exception to the tracking invariant above: it leaves display type at
+`normal` and uses tracking only to open small mono caps. If you take that mode, take it whole.
+
+Pick one. Averaging them produces the stock look with extra steps: the base unit, the grain and
+the radius policy contradict each other directly, and a blend resolves each contradiction toward
+the default.

claude/skills/impeccable/reference/modes/brand.schema.json

@@ -0,0 +1,93 @@
+{
+  "$schema": "http://json-schema.org/draft-07/schema#",
+  "title": ".impeccable/brand.json",
+  "description": "Per-project opt-in. Absent by default: with no brand.json, ui-engineer and component-registry behave normally and ui-gate falls back to its default type scale. Present, it is read by ui-gate/rules/tokens.sh (TOK-TYPE-SCALE) and by impeccable. A skill that fires on whether something deserves to be beautiful will fire wrongly and constantly; a file in the repo is a decision somebody made.",
+  "type": "object",
+  "required": ["mode", "type"],
+  "properties": {
+    "mode": {
+      "enum": ["pixel", "plate", "editorial", "custom"],
+      "description": "Which measured mode this surface targets. See reference/modes/. Not a style preference: the modes contradict each other on base unit, grain and radius policy, and averaging them resolves each contradiction toward the stock default."
+    },
+    "type": {
+      "type": "object",
+      "required": ["scale"],
+      "properties": {
+        "scale": {
+          "type": "array", "items": {"type": "number"}, "minItems": 3,
+          "description": "Font sizes in px. TOK-TYPE-SCALE fails any font-size not on this list. Measured references have no modular ratio: pixel has a 2.31x hole, plate runs two regimes, editorial uses 19 bespoke clamps. State the sizes, not a ratio."
+        },
+        "tracking": {
+          "type": "array",
+          "description": "Tracking as a monotonic function of size, in em. Every reference locks this in em so it survives fluid resize untouched.",
+          "items": {
+            "type": "object",
+            "required": ["minPx", "em"],
+            "properties": {
+              "minPx": {"type": "number"},
+              "em": {"type": "number", "description": "Negative on display, 0 mid-band, positive on small uppercase. Crossover sits at 15-20px in all three references."}
+            }
+          }
+        },
+        "leading": {
+          "type": "object",
+          "description": "Line-height inverts with size in every reference. No exceptions found.",
+          "properties": {
+            "display": {"type": "number", "description": "0.88-0.95 measured"},
+            "mid": {"type": "number", "description": "1.35-1.5 measured"},
+            "body": {"type": "number", "description": "1.5-1.75 measured"}
+          }
+        },
+        "measureCh": {"type": "number", "description": "Max prose line length. 32.6-62 measured; all under 60 except plate's prose at 62. Machine output is exempt and runs long on purpose."}
+      }
+    },
+    "color": {
+      "type": "object",
+      "properties": {
+        "ground": {"type": "string", "description": "Never pure black. Measured: #050505, #09090b, #060708."},
+        "ink": {"type": "string", "description": "Never pure white on dark. Measured: #f7f4ec (4.3% warm), #ecebf2 (cool)."},
+        "accent": {"type": "string", "description": "One hue, or none. pixel has no type accent and reserves #ffffff for a single line."},
+        "hairline": {"type": "string", "description": "Ink at low alpha. plate's most-used value is rgba(236,235,242,0.08) at 1033 occurrences; it has no solid grey borders at all."}
+      }
+    },
+    "spacing": {
+      "type": "object",
+      "properties": {
+        "baseUnit": {"type": "number", "enum": [2, 4, 8], "description": "plate is 2 (85.8% of values), pixel and editorial are 4. Not 8 in any reference."},
+        "ratio": {
+          "type": "array", "items": {"type": "number"},
+          "description": "Tight groups inside enormous silence. Measured: editorial 7:1, plate 1:1.5:3:5:16, pixel 24:1. The ratio is the signature, not any single value."
+        }
+      }
+    },
+    "motion": {
+      "type": "object",
+      "properties": {
+        "hoverMs": {"type": "number", "description": "130-200 in all references."},
+        "entranceMs": {"type": "number", "description": "420-760 in all references."},
+        "staggerMs": {"type": "number", "description": "45 (editorial), 75-80 (plate)."},
+        "easing": {"type": "object", "properties": {
+          "micro": {"type": "string", "description": "cubic-bezier(.4,0,.2,1)"},
+          "entrance": {"type": "string", "description": "cubic-bezier(.16,1,.3,1) or cubic-bezier(0,0,.2,1)"}
+        }}
+      },
+      "description": "Two bands and a hole: nothing between 300 and 420ms in any reference."
+    },
+    "radius": {
+      "type": "array", "items": {"type": "number"},
+      "description": "editorial is binary: 0 or 9999, nothing between, 551 elements against 14. plate runs 8/12/28. Declare the set; anything off it is drift."
+    },
+    "texture": {
+      "type": "object",
+      "description": "Optional. editorial ships none at all and still reads expensive, so absence is a valid choice rather than an omission.",
+      "properties": {
+        "grain": {"type": "object", "properties": {
+          "baseFrequency": {"type": "number", "description": "0.85 (plate), 1.15 (pixel)"},
+          "numOctaves": {"type": "number"},
+          "opacity": {"type": "number", "description": "0.06 overlay (plate), 0.52 screen (pixel)"},
+          "blendMode": {"type": "string"}
+        }}
+      }
+    }
+  }
+}