Build carousel art studio foundation

.claude/skills/carousel-art-studio/SKILL.md

@@ -0,0 +1,169 @@
+---
+name: carousel-art-studio
+description: Art-direct and construct show-stopping, evidence-grounded computational artwork for Texas AI Docket LinkedIn carousels. Use when researching visual references, inventing a carousel treatment, making or comparing styleframes, planning shots, coding bespoke slide art, or reviewing whether a deck reaches the publication's 10x visual ceiling. Produces visual theses, rendered alternatives and production decisions; use carousel-engine separately for rendering and machine QA.
+---
+
+# Carousel art studio
+
+Make an image worth stopping for before making a deck worth approving. This skill is the
+creative studio upstream of `carousel-engine`: it decides what the art says, finds a singular
+visual world, proves that world in pixels and develops the shots. The engine renders and checks
+the result; it does not provide the art direction.
+
+The standard is `knowledge/carousel/VISUAL_CEILING_AUDIT.md`. The source shelf is
+`knowledge/carousel/REFERENCE_ATLAS.md`. The publication rules in
+`knowledge/carousel/DESIGN_DOCTRINE.md` and the accuracy/technical contracts in
+`carousel-engine` still apply.
+
+## Read by task, not by habit
+
+Load only the references needed for the present decision:
+
+- concept or treatment: `references/visual-thesis.md` and
+  `references/composition-and-camera.md`;
+- styleframe competition: `references/styleframe-brief.md` and
+  `references/linkedin-native.md`;
+- dimensional scene: `references/light-material-depth.md`;
+- generative, data-bound or simulated form: `references/computational-systems.md`;
+- full-deck sequencing or pixel review: `references/linkedin-native.md` and
+  `references/review-gates.md`;
+- coding an advanced construction: `references/atelier.md`, then inspect only the relevant
+  source under `examples/art-atelier/`.
+
+Do not read an atelier frame as a template. Read it to understand how camera, value, light,
+material, typography and code resolve into one shot.
+
+## Non-negotiable output
+
+Do not select a treatment from prose. A selected direction requires:
+
+1. a one-sentence visual thesis tied to the verified claim;
+2. a cited reference cut with study/transform/do-not-copy notes;
+3. at least three materially divergent candidate directions;
+4. rendered pixels for each candidate — cover plus one interior turn when selecting a deck
+   language, not a cover alone;
+5. full-size and 432-pixel comparisons;
+6. a written selection decision and discarded-direction reasons;
+7. a shot card for every production slide.
+
+If the task does not permit rendering, label the result **unproven treatment**. Never imply
+that prose has passed the visual gate.
+
+## Studio sequence
+
+### 1. Lock the visual truth
+
+Extract one claim the image can honestly carry. Separate:
+
+- verified fact, quantity, place and time;
+- measured or modeled data;
+- bounded metaphor;
+- unknown or unavailable information.
+
+Geometry is a sentence. Size, position, direction, connection, material and absence all make
+claims even without labels. Do not invent a magnitude, map location, causal link, crowd,
+machine detail or record fragment for drama. Read `references/visual-thesis.md`.
+
+### 2. Cut references with intent
+
+Use three to six references from the atlas and primary story-domain sources. Research a craft
+property, not a vibe. For each reference write what to study, what evidence it serves, how it
+will be transformed and what must not be copied. At least one source must come from the actual
+material world of the story and at least two must break from the current AI-design defaults.
+
+Do not collect a wall of similar images. That narrows invention while pretending to expand it.
+
+### 3. Write the visual thesis
+
+The thesis names a visible event, not an adjective stack:
+
+> The record becomes **[specific world/object/system]**, seen from **[decisive vantage]**, where
+> **[verified tension/change]** is expressed by **[spatial/material behavior]** and illuminated
+> by **[motivated light]**.
+
+Then name the cover's memory sentence: the concrete description a reader could repeat tomorrow.
+If it is “text over a nice background,” restart.
+
+### 4. Diverge in construction, not decoration
+
+Make at least three directions that differ in all of these:
+
+- image architecture: object, environment, section, field, machine, evidence excavation, etc.;
+- vantage and lens: plan, macro, low wide, orthographic, cutaway, impossible evidence camera;
+- material and light behavior;
+- type's spatial role;
+- motion or generative rule;
+- feed-size silhouette.
+
+A palette swap, font swap or alternate arrangement of the same plate is one direction.
+Explicitly challenge the cream/serif archival plate, near-black/bright-accent object plate and
+hairline broadsheet layout. They may win only for a claim-specific reason.
+
+### 5. Prove the hard parts in pixels
+
+For each candidate, build a graybox before finish:
+
+1. three-value composition and crop;
+2. subject silhouette and camera;
+3. depth planes and type reserve;
+4. key/fill/practical light and contact;
+5. material response;
+6. evidence layer and annotations;
+7. finish: atmosphere, grade, grain and optical detail.
+
+Render at 1080×1350 and inspect at 432 pixels. If the graybox is weak, do not texture it.
+Use `references/styleframe-brief.md` for the artifact and `carousel-engine` for rendering.
+
+### 6. Choose against the ceiling
+
+Score candidates with `references/review-gates.md`, beginning with feed stop, story specificity
+and memorability. Accuracy and legibility are vetoes. Select the direction with the strongest
+image system, not the one that was easiest to code or most eloquently described. Preserve the
+losers and the decision so tomorrow does not rediscover them.
+
+### 7. Direct the deck as a sequence
+
+Give every slide a different visual job inside one world. Alternate pressure and release:
+scale, density, camera distance, value, tempo and evidence mode. A quiet frame can be
+exceptional; it still needs a decisive image, proportion or absence. Read
+`references/linkedin-native.md`.
+
+Write a shot card per slide before production. If three cards differ only in copy, the deck is
+still a template.
+
+### 8. Build by passes
+
+Keep construction passes explicit so criticism can target the real problem:
+
+- `G0 truth` — encoded geometry is supported;
+- `G1 read` — graybox and silhouette work at feed size;
+- `G2 space` — camera, overlap, occlusion and scale read;
+- `G3 light` — value hierarchy and motivated illumination reveal form;
+- `G4 material` — surfaces respond differently and specifically;
+- `G5 evidence` — labels, data and source matter integrate honestly;
+- `G6 type` — display and body type participate without becoming furniture;
+- `G7 finish` — atmosphere, grade and microdetail unify the resolved shot.
+
+Do not jump from G0 to G7. Finish amplifies what is already there, including weakness.
+
+### 9. Review the image and the system
+
+Use the full render, the 432-pixel thumbnail and the contact sheet. Review the frame with type
+briefly hidden to expose compositional dependence. Review the deck in grayscale to expose value
+collapse. Then run the machine QA. A green harness means the artwork is safe to judge, not that
+it is great.
+
+## Studio prohibitions
+
+- No invented facts disguised as visual specificity.
+- No adjective-only art direction: premium, cinematic, tactile and bold are outcomes.
+- No named technique standing in for a shot.
+- No full deck before divergent styleframes.
+- No generic background whose only job is filling space behind type.
+- No panel score used as permission for repeatable sameness.
+- No copyrighted style or living artist requested “in the style of.” Study transferable craft
+  properties from original sources and transform them.
+- No external runtime dependency or network asset in slide HTML.
+
+The goal is not maximal visual noise. It is maximal authorship: the most inevitable, surprising
+and beautifully resolved image this particular public record can truthfully support.

.claude/skills/carousel-art-studio/references/atelier.md

@@ -0,0 +1,59 @@
+# Code atelier — reading finished constructions
+
+`examples/art-atelier/` is a set of executable visual studies made with the same offline assets
+as production. Each frame isolates a larger construction — shot, camera, light, material,
+evidence and type — rather than advertising a primitive.
+
+## How to use it
+
+1. Choose the construction whose **problem** resembles the story's, not the one whose look you
+   like most.
+2. Render it through `carousel-engine`; do not judge source code names or a browser at arbitrary
+   zoom.
+3. Inspect the full render and 432-pixel thumbnail.
+4. Read how macro composition is separated from micro detail.
+5. Extract a mechanism: depth ordering, highlight construction, type reserve, data-to-form
+   mapping, evidence integration.
+6. Rebuild the mechanism with the day's own nouns, camera, palette, material and behavior.
+
+Never duplicate an atelier frame and replace its text. That would turn a craft reference into
+the template this skill exists to prevent.
+
+## Construction index
+
+| Frame | Problem it demonstrates | Inspect for |
+|---|---|---|
+| 01 — evidence monolith | Make a record feel physical and consequential | Extreme crop, lit edge, embedded evidence, type as counterforce |
+| 02 — volumetric cutaway | Reveal concealed layers/process without a dashboard | Impossible evidence camera, depth separation, motivated internal light |
+| 03 — topographic signal | Turn a verified gradient or threshold into a material field | Macro density transition, contour survival, one literal anchor |
+| 04 — data sculpture | Give a small verified series scale and spatial presence | Bound geometry, camera, lit ground, quantitative legend discipline |
+| 05 — document archaeology | Use real record matter without making a paper collage | Layered depth, raking light, selective focus, exact fragment provenance |
+| 06 — typographic machine | Let language participate in physical behavior | Type as structure, moving parts, legibility protected by construction |
+
+These frames are technique-agnostic at the concept level. Production may implement the same
+construction with Canvas, SVG, relief, SDF or Three.js depending on the shot budget.
+
+## What to copy and what not to copy
+
+Copy:
+
+- explicit value hierarchy;
+- camera and depth planning;
+- motivated light;
+- material differentiation;
+- a macro read plus micro reward;
+- deterministic mechanics;
+- protected vector type;
+- comments explaining visual decisions.
+
+Do not copy:
+
+- palette;
+- type lockup;
+- object silhouette;
+- exact shader/noise parameters;
+- frame layout;
+- story metaphor.
+
+The atelier is successful when tomorrow's work looks nothing like it but is built with the same
+degree of intention.

.claude/skills/carousel-art-studio/references/composition-and-camera.md

@@ -0,0 +1,86 @@
+# Composition and camera
+
+## Compose the thumbnail first
+
+At 432 pixels wide, detail collapses before structure. Begin with three flat values and ask:
+
+- Is the main silhouette recognizable with type hidden?
+- Is one region dominant, one subordinate and one quiet?
+- Does the eye enter at a deliberate point?
+- Is the crop active, or is the subject politely centered?
+- Does negative space create pressure, scale or anticipation?
+
+Make six postage-stamp grayboxes before polishing one. A successful graybox should still carry
+the visual thesis in grayscale.
+
+## Name the camera
+
+Every dimensional frame needs a camera note:
+
+```text
+Vantage:
+Lens / projection:
+Camera height and tilt:
+Target:
+Horizon:
+Near / middle / far planes:
+Crop and intended offscreen continuation:
+Why this camera tells the claim:
+```
+
+Useful choices have meaning:
+
+- low wide: monumentality, exposure, instability, distance;
+- long lens: compression, congestion, stacked relation;
+- macro: material evidence, hidden mechanism, fragility;
+- orthographic: authority, comparison, system legibility;
+- plan: territory, allocation, pattern;
+- cutaway: concealed structure or process;
+- tilted/oblique: tension, transition, competing planes.
+
+Do not pick a dramatic lens because it looks dramatic. Pick it because perspective changes the
+reader's understanding.
+
+## Build depth before texture
+
+Use at least three legible spatial layers where depth matters:
+
+1. foreground occluder or scale cue;
+2. subject plane;
+3. ground/environment;
+4. atmosphere or distant field when useful.
+
+Separate layers by overlap, scale, focus, value, haze and motion direction. A stack of same-
+value translucent rectangles is layering, not depth.
+
+## Type has a spatial role
+
+Choose one, deliberately:
+
+- **monument:** type is a massive form the image negotiates;
+- **aperture:** type or its reserve reveals/conceals the world;
+- **instrument:** type belongs to a measurement or machine layer;
+- **caption:** quiet type observes an image that already works;
+- **counterforce:** type pushes against the subject mass;
+- **evidence:** exact record language is materially embedded.
+
+Avoid placing a self-contained text box into leftover space. Type and art can remain separate
+for legibility while still sharing proportion, axis, scale or rhythm.
+
+## Deck rhythm
+
+Map the sequence before coding:
+
+| Beat | Typical job | Spatial move |
+|---|---|---|
+| Stop | Singular hook | Extreme silhouette/crop, shortest copy |
+| Orient | Establish world | Wider camera, literal anchor |
+| Reveal | Expose mechanism | Cutaway, macro or layer separation |
+| Measure | Make evidence felt | Data-bound form, comparison or scale shift |
+| Turn | Change understanding | Reversal of value, vantage or behavior |
+| Consequence | Human/institutional meaning | Specific object, place or absence |
+| Resolve | Useful takeaway | Clearer space, decisive final image |
+
+Not every deck needs seven beats and a nine-slide deck may split them. It does need contrast.
+Alternate full/empty, near/far, dense/quiet, orthographic/perspectival and dark/light without
+abandoning the same world.

.claude/skills/carousel-art-studio/references/computational-systems.md

@@ -0,0 +1,68 @@
+# Computational systems — behavior with meaning
+
+## Pick a behavior, not an effect
+
+Describe the system as a sentence:
+
+> **[entities]** move/grow/partition/accumulate according to **[rule]**; **[verified input]**
+> controls **[visible property]**; the viewer should perceive **[claim]**.
+
+Examples of meaningful mappings:
+
+- delay controls queue length, not arbitrary hue;
+- uncertainty controls blur or boundary instability, not fake precision;
+- load controls density or deformation when the scale is documented;
+- jurisdiction controls real polygon membership, not invented cells;
+- time controls position or erosion along one declared direction.
+
+If there is no verified input, label the system illustrative and make no legend that suggests
+measurement.
+
+## System families
+
+| Behavior | Good for | Common lie/failure |
+|---|---|---|
+| Advection / flow | movement, transmission, drift | random streamlines imply routes or volume |
+| Accumulation | queue, backlog, compounding | area looks quantitative without a scale |
+| Field / contour | threshold, gradient, influence | decorative noise masquerades as measured terrain |
+| Packing / collision | capacity, congestion, constraint | circle size implies magnitude accidentally |
+| Growth / branching | propagation, dependency, buildout | invented topology reads as a real network |
+| Partition | allocation, territory, ownership | Voronoi cells replace real jurisdictions |
+| Erosion / fracture | loss, stress, exposure | damage metaphor overstates an administrative change |
+| Oscillation / wave | cycles, instability, signal | amplitude or frequency appears data-bound when it is not |
+
+## Authored generative work
+
+Randomness is not art direction. Define:
+
+- invariant structure — what every seed must preserve;
+- bounded variation — which properties may move and within what range;
+- composition constraints — quiet zone, dominant mass, collision rules;
+- selection policy — whether a seed is fixed before or after styleframe review;
+- failure detection — empty field, even density, hair-like lines, moiré, bad crop.
+
+Use deterministic seeds. Never sample seeds until one “looks cool” without recording the
+selection, because that hides art direction inside roulette.
+
+## Data, model and metaphor labels
+
+Classify every visual layer:
+
+- **DATA** — directly bound to a cited value;
+- **MODEL** — produced by a documented transformation or simulation;
+- **MAP** — real geographic geometry/projection;
+- **ILLUSTRATION** — explanatory but not measurable;
+- **ATMOSPHERE** — non-semantic finish.
+
+The dossier should say which is which. Data may drive illustration, but the reader must not be
+invited to recover a number the mapping cannot support.
+
+## Thumbnail stability
+
+Generative detail dies first at 432 pixels. Design two scales:
+
+- macro structure: one silhouette, density transition or directional field;
+- micro reward: marks, particles, hachure, interference or grain.
+
+If removing the micro layer erases the claim, the system is under-composed. Avoid line spacing
+and repetition frequencies that moiré after downsampling; render the actual thumbnail.