Imported from TaiT-tt/tait-crt-interface-skill (SKILL.md). Install upstream with npx skills add TaiT-tt/tait-crt-interface-skill. Copyright stays with the author.

tait-crt-interface-skill

Mode Policy

Use one shared content pipeline with two capability-gated execution profiles. Resolve the profile once, parse any optional named style preset, resolve palette, resolve ratio, inspect once, anchor the subject roster, build one abstraction blueprint, apply the matching preset recipe or select one normal variation recipe, compile one internal prompt, generate once, complete only the selected profile's delivery path, inspect once, and return one image.

Execution Profile Gate

Resolve execution_profile from advertised capabilities before intake. Use capabilities—not product names, self-description, or user choice—and never call an unavailable tool merely to test it. Retain the profile across palette and ratio turns; never ask the user to choose a profile or switch profiles after generation starts.

• Select codex-full only when the environment exposes the exact callable codex_app__load_workspace_dependencies, local script execution, and a readable local path for the generated image. This profile preserves the existing dependency lookup, Pillow probe, same-resolution scripts/finalize_crt.py pass, report inspection, and finalized PNG output exactly as written below.
• Otherwise select portable-direct when an image-generation capability exists. Never call codex_app__load_workspace_dependencies, inspect Python or Pillow, search for interpreters, run local scripts, reopen or convert the generated image, or attempt any post-processing. Compile the finalizer-owned visual requirements into the single generation prompt and return that generated image directly after one visual inspection.
• If no image-generation capability exists, stop once with a concise unsupported-environment message. Do not attempt dependency discovery, script execution, prompt-only text output, or repeated recovery.

Two-Stage Intake Gate

Before any image-generation call, parse the current request and retained conversation state for palette and aspect_ratio. Read references/palettes.md whenever selecting, displaying, deriving, or validating color.

1. If palette is missing, in codex-full display the corrected bundled card assets/tait-crt-interface-color-card.png and use the compact text list in references/palettes.md only if the image cannot be rendered; in portable-direct, use that compact text list immediately without resolving or rendering the asset. Ask the user to reply with one card name or 如图, then stop the turn. Do not ask for ratio in the same turn and do not generate.
2. After palette is known, if aspect_ratio is missing, ask only: 请选择生成比例：3:4 / 4:3 / 9:16 / 16:9，也可以直接输入其他比例。 Then stop the turn and do not generate.
3. Generate immediately only when both fields are known. If the initial invocation supplies one field, skip only that question; if it supplies both, skip the entire intake. Do not reconfirm valid choices.

Recognize 经典, 粉黛, 极客01, 极客02, 复古01, 复古02, 游戏01, 游戏02, and 如图 as palette values. For 如图, derive a coherent 2-5-color palette from the uploaded reference's visual mood, atmosphere, and emotional tone; use no more than five hues and proceed without another aesthetic questionnaire. If no usable image is present, request the missing image as the sole blocking exception.

Optional Style Preset Router

Activate style_preset only when the user explicitly selects one exact value—街头怪诞, 巨像符号, or 冷面几何—as the requested style for the current generation, such as “用街头怪诞生成” or by listing it with the palette and ratio. Quoted, negated, compared, documented, or merely discussed names do not activate or persist. Once explicitly selected for a generation, retain it across palette/ratio intake turns and read references/style-presets.md before building the blueprint. Merge its fixed style block with the uploaded subject's own roster, identity anchors, actions, and objects; never copy example props or invent missing traits merely to resemble a preset example.

The preset locks only its listed subject-shaping axes. Select all remaining interface/layout axes from the normal Variation Engine so the result still fits the reference, subject count, and selected ratio. Preserve the locked style axes unless crop or coverage must expand to keep every roster member and interaction; never weaken invariants. If no preset was explicitly selected for the current generation, do not mention, offer, recommend, or ask about presets, do not load one implicitly, and continue with the original abstraction blueprint plus normal Variation Engine. Never mix presets in one image; if the user explicitly replaces a retained preset with another exact name, use the latest explicit name.

Non-Negotiable Invariants

• Generate exactly one final image at the selected ratio and the tool's native dimensions. Never make a low-resolution precursor, request 480x720, resize, enhance, produce variants, retry automatically, or expose the prompt.
• Treat uploads only as sources of a few identity anchors, never as cutout masks, tracing templates, contour maps, detail inventories, or sources for automatic pixelation. Make the independently authored bitmap subject the first visual priority: preserve recognition through a small anchor set, then redesign everything else. Caricature, abstraction, adult editorial personality, and retro illustration quality outrank literal fidelity. Never preserve source-matching garment construction, decorative motifs, jewelry inventories, folds, textures, facial modeling, or local contours unless one item is an explicit identity anchor. In a multi-subject reference, preserve every anchored figure as a distinct identity; never omit one, fuse bodies or faces, swap identities, or transfer traits between figures.
• Draw exactly one full primary wallpaper composition—one figure or one anchored group—directly on the base field behind every window. Its silhouette covers at least 50% of the canvas; target 60%-80%. Never repeat it. A group remains one composition while retaining every anchored member, their spatial order, and their interaction. Include 1-3 feature-extraction windows showing only distinctive partial elements—not another full subject or group.
• Preserve 20%-30% connected open field while using 3-6 foreground windows with clear size hierarchy, staggered overlap, multi-quadrant balance, a full-width menu, one open French drop-down, and exactly one cursor.
• Use only the resolved palette. Named palettes use their listed colors; 如图 uses its derived 2-5 colors. Prefer bold adjacent flat-color planes. Use a regular darkest/lightest checkerboard only when a non-skin curved surface needs one optical middle tone or when touching planes cannot otherwise communicate front/back structure. Checkerboard is optional, never decorative, and forbidden on faces, necks, ears, hands, arms, legs, or any exposed skin.
• Establish one global content lattice with square base cell p, approximately short canvas edge / 384. Snap the subject, windows, borders, glyphs, icons, cursor, charts, accents, and checkerboard to this same p x p cell; never give the subject a separate coarser pixel size. Treat scanlines/noise as CRT surface texture, not a second content grid.
• Apply hard square pixels, dense scanlines, palette-bound bloom, noise, persistence, misregistration, one restrained sync fault, and unmistakable spherical barrel curvature/compression in the outer 10% of all four edges while keeping the inner 80% stable. This distortion is mandatory, never a recipe option.
• Reserve the upper-right title-bar zone for the immutable exact lowercase signature tait-crt-interface-skill, rendered in the same global bitmap grid and family as the left label rail. Never omit, translate, rename, crop, cover, hide, or deliberately corrupt it, even if a later request asks.

Standard Prompt Compiler

Compile only requirements that become pixels. Answer these fields in order.

1. Abstraction Blueprint

Treat the subject illustration as priority 1. When literal likeness conflicts with a strong authored cartoon, keep only the identity anchors and choose the stronger cartoon. Record internally:

First run Subject Recognition and Anchoring. Identify all visually prominent or interacting people, creatures, or objects that plausibly form the intended subject. Lock a roster as S1, S2, and so on, plus left/right/front/back order, relative scale, pose/action relationship, contact and occlusion, and held or shared objects. Treat distant incidental bystanders as background, but when a prominent or interacting figure is ambiguous, preserve it rather than silently deleting it. Unless the user explicitly selects one figure, the roster count is immutable: never merge two figures into one hybrid or borrow one figure's face, hair, clothing, limbs, or accessories for another.

When style_preset is active, bind the matching fixed style block after roster locking and before the numbered blueprint fields. Translate its roles onto the actual subject—human, group, creature, object, or mixed roster—using only observed identity anchors. The preset controls how the subject is redesigned, never which subject content exists. When no preset is active, derive all blueprint choices normally.

1. identity anchors: keep only 3-5 high-information traits for each rostered subject and bind every anchor to its subject ID. Prefer one dominant silhouette or hair-mass trait, one facial-spacing or head-shape trait, one pose/action relation, and at most one signature accessory or object. An anchor is a semantic cue, not permission to reproduce its exact contour or decoration. Do not inventory clothing panels, prints, embroidery, jewelry, folds, wrinkles, skin texture, lighting, or anatomical detail.
2. source severance: convert the anchors into a short text-only character brief, then rebuild from that brief as if the photograph were unavailable. Deliberately abandon the original cutout contour, crop logic, facial map, body proportions, modeled volume, shadow pattern, garment construction, decorative placement, and photographic edge continuity. Reusing the source silhouette and merely simplifying its pixels fails; the new subject must be impossible to obtain by masking, tracing, posterizing, or mosaicking the source.
3. caricature mutation: select one identity anchor as the dominant visual hook and exaggerate 2-3 traits around it. Make at least one major proportion mutation and one major silhouette mutation, plus a third change among facial spacing, limb thickness, pose angle, object scale, negative-space rhythm, or accessory scale. Push these changes far enough to feel authored, asymmetrical, awkward, blunt, strange, or charmingly ugly; loose recognition comes only from the anchor set, not from preserving the source anatomy.
4. mass plan: redraw each subject from 5-9 flat interlocking masses, preferably 5-7, with one unmistakable dominant mass, one counter-mass, and a few supporting blocks. Reduce the head and face to a coarse mask of a few stepped clusters; reduce hair to one or two masses, each garment to one to three broad planes, and limbs to blunt structural shapes. Replace volume primarily with contrasting flat colors, cutouts, negative space, straight-edged facets, and hard silhouette breaks. Reserve checkerboard for an indispensable non-skin curved surface or an ambiguous contact between planes. Delete every non-anchor print, ornament, seam, fold, strand, pore, wrinkle, highlight, fingernail, subtle expression, and decorative micro-feature instead of translating it into pixels.
5. authored style lock: render the subject as an adult 1980s bitmap editorial caricature or underground terminal illustration—angular, industrial, graphic, slightly grotesque, and visibly designed from scratch. Favor uneven visual weight, purposeful asymmetry, compressed anatomy, abrupt stepped contours, mask-like facial construction, expressive negative space, low facet counts, straight runs, chamfers, wedges, rectangles, and blunt polygons. Treat curves as high-risk detail channels: default to no curves; when the subject becomes unreadable without one, allow only one or two large stepped arcs built from a few straight segments. Forbid flowing contours, repeated rounded features, realistic pixel portraiture, polished game-sprite anatomy, anime/chibi proportions, kawaii mascot language, round friendly children's-book shapes, decorative fashion illustration, clean vector-cartoon outlines, and generic schoolbook or clip-art character design.
6. hand topology: for every visible hand, bind left/right ownership, wrist attachment, palm orientation, contact or occlusion, and any clearly visible finger count before simplification. Use a readable blocky palm with anatomically plausible finger grouping. If the grid cannot support separate fingers, use an intentional mitten or grouped silhouette with no false finger tips; never create extra or missing fingers, fused hands, detached palms, inverted joints, impossible wrist connections, or hand-object penetration.
7. pixel topology: choose one global square base cell p, approximately short canvas edge / 384, for the entire interface. Use an approximately 30 x 30 planning grid only as the subject's contour/feature complexity budget, then build every subject shape as complete integer-aligned runs of the same p x p cells used by UI and text. Construct with horizontal/vertical runs, deliberate stair diagonals, chamfers, wedges, and low-sided polygons. Use a curve-free geometry recipe whenever possible; any indispensable curve becomes one coarse stepped arc assembled from 3-6 straight segments, never a smooth or continuously varying contour. Permit no half cells, stretched or rotated cells, sub-cell fragments, shifted rows, broken checker patterns, accidental cracks, isolated debris, mixed pixel scales, or smooth transition pixels. Do not render or resize an intermediate image.

Before prompt compilation, run eight rejection tests once. Roster integrity: every anchored subject remains distinct, correctly ordered, and recognizable through its own anchors. Hand integrity: every visible arm reaches the correct wrist and hand, and every depicted finger grouping is plausible. Silhouette divergence: the rebuilt subject must not align closely with the photograph's continuous outer contour, proportions, and pose. Detail amputation: no source garment construction, pattern, ornament, fold, jewelry inventory, facial modeling, or local contour survives unless it is the single declared signature anchor. Filter removal: after mentally removing UI and CRT texture, the subject must still read as a separately authored flat caricature, never a photograph underneath pixelation. Style independence: the subject must read as an angular adult bitmap editorial illustration, not realistic pixel portraiture, a polished game sprite, anime/chibi art, cute mascot art, children's-book illustration, or generic clip art. Skin cleanliness: every face and exposed-skin region is built only from solid palette planes with no checkerboard, stipple, hatch, or spotted texture. Curve austerity: geometry is predominantly straight, faceted, and low-sided; no recipe contains flowing contours or a chain of small variable arcs, and any indispensable stepped arc is large, sparse, and structurally necessary. If a test fails, correct the blueprint once by deleting fidelity, checker texture, or curves and increasing flat geometric redesign instead of adding texture, detail, or another generation loop.

Assign broad tonal planes with bold solid palette-color collisions. Add a regular 50% darkest/lightest checkerboard only to the smallest necessary non-skin region: a curved object needing one optical middle tone or touching planes whose depth relation is otherwise unreadable. Zero checkerboard is valid. Never place it on a face or exposed skin, and never use it as all-over texture. Never rely on realistic rendering, cutout, pixel filter, auto-pixelate, blur, antialiasing, gradient transitions, smooth curves, or photographic edge continuity.

2. Attention and Interface Geometry

• Place the single wallpaper composition—one figure or one anchored group—off-center using the recipe; its final silhouette occupies 60%-80% when possible and never below 50%, including after window occlusion.
• Give windows obvious large/medium/small tiers; the dominant is about 1.8x a medium and the smallest no more than half a medium. Spread centers across at least two quadrants, preferably three, and never stack three similar windows on one side.
• Overlap selected windows by 5%-20%, cross the subject boundary, and keep overall visual mass balanced. Never frame the primary subject or place it in a portrait/bitmap-viewer window.
• Use square borders, title bars, close boxes, scroll tracks, primitive tables/charts, short French bitmap labels, the top menu/drop-down, and one cursor.
• Include 1-3 magnifier-like feature-extraction windows. Isolate recognizable local elements and show each as a tight crop. Magnify by allocating more shared-grid cells to the crop, never by enlarging its cell beyond p. With multiple extractions, vary both outer dimensions and aspect ratios; no two share the same size or proportions.

3. Color and Pixel Construction

• Apply the resolved palette only. Use its lightest and darkest colors for field/ink contrast and checkerboard endpoints; use remaining colors as intentional flat planes, title bars, panels, icons, or signal accents. The selected polarity recipe determines whether the dominant field is light, dark, or locally split; no polarity is the universal default.
• Permit no color outside the palette, alpha, intermediate computed gray, gradient, translucent glow, or antialiasing. Prefer separate solid palette planes; when a true optical half-tone is structurally necessary, use one compact regular darkest/lightest checkerboard region. Keep every face and exposed-skin region solid and checker-free. Do not scatter checks as noise, use them decoratively, or cover the figure with them.
• Build subject edges, UI geometry, icons, and 5x7, 6x8, or 7x9 glyphs from the same global square cells and coarse straight runs, stair diagonals, chamfers, wedges, and low-sided polygons. Simulate bloom only as palette-bound hard-cell expansion; do not use checker fringe on skin. Curves are disabled by default and limited to one or two indispensable coarse stepped arcs in any recipe that permits them.

4. CRT Signal Surface

Cover the frame with dense palette-bound scanlines, sparse noise, one-cell misregistration, short vertical persistence, and one restrained sync disturbance. Force radial barrel distortion inside the outer 10% on every side with nearest-cell displacement: corners compress more than edge centers, the top menu baseline and side rails visibly bow, and no finished frame remains rectangular or flat. Keep the inner 80% stable and show no physical monitor, bezel, or room.

Prompt Shape

Write four compact internal paragraphs:

1. selected ratio, active named style preset and its fixed style block or no preset, locked subject roster and relationships, 3-5 semantic identity anchors per subject, discarded literal details, source severance, major proportion and silhouette mutations, selected adult bitmap editorial style lock, selected low-facet geometry grammar and curve budget, 5-9-mass CRT-caricature blueprint, hand topology, approximately 30x30 contour-detail budget on the global lattice, one wallpaper composition, crop, position, and coverage
2. window count, constellation, hierarchy, overlap, 1-3 unequal feature extractions, menu, labels, cursor, and reserved signature zone
3. selected or derived palette, polarity recipe, conditional non-skin checkerboard placement or none, solid-color skin planes, and one shared subject/UI/text lattice
4. CRT signal/edge warp, industrial mood, and hard avoids

Prefer decisive geometry and measurable shares over a style essay. Never reveal this prompt unless requested.

When execution_profile is portable-direct, make paragraph 4 a direct-render closure: state that the generator's one image is already the final artifact and must itself contain the resolved palette, whole-cell shared lattice, hard bitmap glyphs, palette-bound scanlines/bloom/noise/misregistration/persistence, outer-10% spherical barrel curvature, and exact upper-right signature. Do not mention scripts, Python, Pillow, post-processing, reports, or a later correction pass in the prompt.

Variation Engine

Before compiling, choose one option from every axis. User constraints override recipe choices but never invariants. When style_preset is active, copy its six locked subject-axis values from references/style-presets.md and choose only the remaining axes here; otherwise choose every axis normally. Variation changes visual grammar—not only position. If recent visible outputs repeat a layout, change at least two unlocked high-impact axes among placement, crop, window constellation, count, and hierarchy.

Compatibility rules:

• Pair 80% subject coverage with 3-4 windows; pair 6 windows with 60%-70% coverage and smaller utilities so open field and subject dominance survive.
• Use 1-3 extraction windows and never show a whole face, body, creature, or object twice. With two or three, use different features, sizes, and aspect ratios. Count them within the 3-6 total windows: one extraction permits 3-6 total, two require 4-6, and three require 5-6.
• Default midtone map to none; select another option only when flat palette-color collisions cannot express curvature or front/back structure. Keep the checkerboard compact, connected, nondecorative, and entirely outside faces and exposed skin. Resolve ordinary plane differences with contrasting palette colors instead.
• Pair the first three geometry grammar options freely. Use limited-single-stepped-arc only for an identity-critical rounded silhouette or indispensable spherical object, and cap it at one or two large arcs made from 3-6 straight segments. Never use several variable curves to recover source detail.
• Every recipe must pair one cartoon treatment with one caricature mutation; neither may preserve source-matching anatomy, garment detail, or local contours. If a treatment drifts toward realism or children's illustration, replace it with angular-terminal-caricature before compilation.
• A named preset's fixed style block and locked subject-axis recipe take precedence over normal recipe selection. Do not blend preset language or axes with another preset. Unlocked axes remain variable and must adapt to the selected ratio and actual roster.
• Rebalance or simplify windows before reducing the subject below 50%, removing open field, or weakening the selected palette.

Workflow

1. Resolve and retain execution_profile, parse and retain any explicitly selected exact optional style_preset, then complete the Two-Stage Intake Gate. Never ask about profiles or presets. Do not inspect for generation or call image generation until both palette and ratio are known.

2. Inspect the reference once; run Subject Recognition and Anchoring, lock the subject roster and relationships, extract the abstraction blueprint, and, for 如图, derive the 2-5-color palette.

3. If style_preset is active, bind its fixed style block and locked six-axis subject recipe, then select the remaining axes once. Otherwise select one complete normal recipe. Do not discuss alternatives or default repeatedly to the same layout.

4. Preflight once: confirm the locked subject count and relations, per-subject identity binding with no omitted or hybrid figures, plausible hand ownership/wrist/palm/finger topology, source severance, only 3-5 semantic identity anchors per subject, no copied clothing/ornament/fold/local-contour detail, one major proportion mutation, one major silhouette mutation, at least one additional structural change, a 5-9-mass adult bitmap editorial caricature that passes all eight rejection tests, approximately 30 x 30 contour-detail budget, complete square integer-aligned global cells with no broken topology or smooth sub-cell transitions, a selected low-facet geometry grammar with curves disabled or tightly budgeted, solid checker-free faces and exposed skin, checkerboard set to none or one justified compact non-skin structural region, one borderless wallpaper composition, 60%-80% target coverage, 1-3 unequal feature extractions, selected ratio/palette, a clear upper-right signature zone, and no realistic/filter/tracing/children's-illustration language.

5. Compile the prompt once, adding the direct-render closure only for portable-direct, and call the available image-generation capability exactly once at native resolution.

6. Complete the selected delivery path:

   • In codex-full, before finalization call codex_app__load_workspace_dependencies once and use the returned bundled Python executable, which includes Pillow. Never invoke bare python, bare python3, or /usr/bin/python3 for this script. Run without resizing, passing the resolved palette and recipe polarity:

     <bundled-python-executable> scripts/finalize_crt.py --input <generated-image> --output <new-final.png> --palette <comma-separated-hex-colors> --polarity <positive|negative|preserve> --grid-cell 0 --edge-warp 0.12

     Map light-field to positive, dark-field to negative, and split-local-fields to preserve. Resolve the script relative to this file, verify the selected interpreter with from PIL import Image before the run, and never overwrite the source. If the import probe fails, select another reported workspace interpreter rather than installing packages or falling back to system Python.

   • In portable-direct, treat the generated image as final. Do not call dependency tools, probe or install packages, run any Python, locate the finalizer, create an intermediate image, or initiate a second generation/editing pass.

7. Inspect once against the profile-specific Quality Gate. Never regenerate automatically. In codex-full return only the finalized image; in portable-direct return only the original generated image.

Quality Gate

• execution_profile was resolved before intake and remained stable. codex-full ran the complete existing finalization path; portable-direct made no dependency, interpreter, Pillow, filesystem-finalizer, image-conversion, or post-processing attempt.
• Both intake fields were resolved before generation; the canvas matches the selected ratio and the final palette matches the named or derived choice.
• If an exact style preset was supplied, the compiled blueprint contains its fixed style block and all six locked subject-axis values while the remaining axes adapt to the actual reference and ratio. If none was supplied, no preset was mentioned or silently applied and the normal blueprint/Variation Engine path was used.
• Priority 1 passes before interface polish: the subject reads immediately as an independently authored, adult 1980s bitmap editorial caricature—angular, graphic, asymmetrical, personality-rich, and optionally strange, funny, awkward, grotesque, or charmingly ugly. Only 3-5 semantic identity anchors survive per roster member; one major proportion mutation, one major silhouette mutation, and at least one additional structural change are plainly visible. Five to nine large masses, not copied local detail, carry recognition. No non-anchor garment construction, print, ornament, jewelry inventory, fold, facial modeling, or photographic contour remains. Its silhouette cannot be closely overlaid on the source, and removing UI/CRT texture cannot reveal realistic pixel portraiture, a traced or filtered photograph, polished game-sprite anatomy, anime/chibi art, cute mascot art, children's-book illustration, or generic clip art.
• The locked roster is intact: every prominent or interacting source subject appears once as a distinct figure in the single wallpaper composition, with correct spatial order, scale relationship, interaction, and subject-bound traits. No figure is omitted, fused, swapped, or transformed into a hybrid. Every visible hand belongs to the correct arm, connects through a plausible wrist and palm, respects contact/occlusion, and has either a correct visible finger count or an intentional undivided grouped silhouette—never malformed pseudo-fingers.
• Subject complexity reads like an approximately 30 x 30 planning grid, but every displayed pixel uses the same complete, square, integer-aligned p x p base cell as window borders, icons, and bitmap glyphs. Inspect subject contours, any checkerboard plane, and UI/text in multiple regions: stair steps are intentional, every content mark resolves to whole square cells, and no subject pixel is several times larger than a UI/text pixel. Geometry is dominated by straight runs, stair diagonals, chamfers, wedges, and low-sided polygons. Any flowing contour, unnecessary sequence of variable arcs, half-cell, stretched or rectangular cell, shifted row, broken checker sequence, accidental gap, misaligned island, stray debris, antialiased edge, or smooth transition pixel fails.
• Tonal depth relies first on bold solid palette planes. Checkerboard is optional; when present, it is one or a few compact connected structural regions made only from the darkest and lightest colors, limited to necessary non-skin curved surfaces or plane-separation cues. Every face, neck, ear, hand, arm, leg, and other exposed-skin area remains solid and checker-free. Decorative, scattered, or all-over checker texture fails.
• Exactly one borderless wallpaper composition covers at least 50%, preferably 60%-80%; it contains the full locked roster, no similarly sized copy appears, and any detail crop is small and partial.
• One to three feature-extraction windows appear. Every crop isolates a distinctive partial element, uses the shared cell p, and differs from every other extraction in both dimensions and aspect ratio.
• Windows follow the selected count, constellation, and hierarchy; they span multiple quadrants, overlap without hiding too much subject, avoid rigid equal-size stacks, and preserve 20%-30% connected open field.
• In codex-full, dimensions match the generated source and the report contains only colors from the resolved 2-5-color palette, confirms no_synthetic_checkerboard, and reports one global grid cell shared by all content. Zero checkerboard is valid because finalization must preserve authored structural checks without creating new ones.
• In portable-direct, judge palette restriction, whole square cells, shared subject/UI/text scale, scanlines, barrel curvature, and the fixed signature visually in the one generated image. Do not claim deterministic pixel, color, resolution, or report verification; do not repair a failure with tools, scripts, resizing, or another image call.
• Bitmap glyphs, palette-bound bloom, scanlines, noise, persistence, and misregistration remain visible. The outer 10% shows unmistakable radial barrel curvature on all four sides, with stronger corner compression and visibly bowed long lines, while the center remains stable.
• The upper-right title bar contains the exact readable lowercase string tait-crt-interface-skill in the same shared-grid bitmap style as the left label rail. It is not missing, misspelled, translated, obscured, or replaced.
• Exactly one final image is returned; no prompt, recipe, checklist, drafts, or alternatives are shown unless requested.

Hard Avoids

• selecting an execution profile from an Agent or product name, asking the user to choose a profile, probing unavailable tools, calling Codex-only tools in portable-direct, inspecting or installing Python/Pillow in portable-direct, running the finalizer in portable-direct, or switching profiles after generation begins
• generation before both intake fields are known, asking palette and ratio in one turn when both are absent, reconfirming supplied fields, or silently substituting a default palette/ratio
• asking users to choose a style preset, recommending one without request, silently defaulting to one, ignoring an exact supplied preset, blending two presets, replacing subject content with preset-example props, or allowing a preset to omit/merge roster members
• realistic likeness painting or pixel portraiture, source-matching silhouette/proportions/pose, preserved garment construction, copied prints/ornaments/jewelry/folds/local contours, cutout plus pixel filter, mosaic overlay, posterization, auto-pixelation, literal tracing, photographic outlines, polished game-sprite anatomy, anime/chibi or kawaii language, cute mascots, children's-book or schoolbook illustration, generic clip art, clean vector-cartoon outlines, fine hair/pores/folds, smooth shading, flowing contours, repeated rounded features, chains of small variable arcs, ornamental curves, blurred or rectangular pixels, subject-only macro pixels larger than UI pixels, half/stretched/rotated cells, smooth sub-cell transitions, shifted rows, broken cell topology, unaligned grids, excessive contour decisions beyond the approximate 30x30 planning budget, or subdivided cells
• omitted, fused, swapped, or hybridized subjects; traits transferred between people; detached or fused hands; impossible wrists, palms, joints, contacts, or occlusions; extra fingers, missing fingers when separate fingers are depicted, or malformed pseudo-fingers; subject below 50%; framed primary composition; duplicate/similarly sized composition; centered symmetry; one-sided window weight; equal-size windows; or three similar windows in one row/column
• colors outside the resolved palette, more than five derived hues, computed gray pixels, checkerboard on any face or exposed skin, decorative/scattered/all-over checker texture, synthetic checkerboard added during finalization, alpha, gradients, soft antialiased glow, modern cards, glassmorphism, contemporary icons, clean vector export, 3D, or game HUD
• missing/flat barrel distortion, identical extraction windows, whole-subject extraction, hidden or altered fixed signature, heavy central distortion, physical monitor housing, logos other than the fixed signature, CTA, long text, low-resolution precursor, resizing, enhancement, retry, variants, or displayed prompt

Maintenance Integrity

Only when modifying this skill, first read references/requirements.md, references/style-presets.md when present, the current SKILL.md, scripts, metadata, and available conversation history. Preserve every existing requirement unless the user explicitly removes or replaces it.

Output

During intake, return only the single required question. After both fields are resolved, return only one image at the generator's original resolution: the finalized PNG in codex-full, or the untouched generator output in portable-direct. If no image-generation capability exists, return only the concise unsupported-environment message.