mixmatter · diff
git:20260918.1616335 to git:20260930.7c220a9
318 added, 242 removed. Audit A to A.
---
name: mixmatter
description: >-
- Art-direct and reconstruct a user-supplied image with MixMatter 2.1: a source-aware contemporary editorial print
- system that reads semantic identity, visual hierarchy, structural relationships, and transformation opportunities,
- then rebuilds the image through selective photography, flattened graphic fields, controlled halftone/duotone,
- source-derived cropping, tactile collage, and modernist hierarchy. For clear requests, execute directly. For vague
- requests, inspect the source and choose the strongest source-specific direction unless a real ambiguity materially
- affects the result. For revisions, preserve successful decisions and change only the requested or diagnosed axis.
- Add zero new text by default unless the user explicitly supplies exact wording or explicitly authorizes generated
- typography. Never translate or bilingual-duplicate source text by default. v2.1.0.
+ Transform a user-supplied photograph into a MixMatter image: a source-aware contemporary editorial print
+ reconstruction using selective photography, visible halftone or duotone treatment, strongly flattened graphic
+ fields, tactile torn-paper collage, modernist hierarchy, and print texture. Use when the user asks for
+ MixMatter, editorial print reconstruction, halftone collage, print-modernist image transformation, or wants
+ a photo rebuilt into a bold non-photographic printed composition. Preserve semantic identity, source-defining
+ structural anchors, and source text when it materially contributes to scene identity. By default, add zero new
+ text. If the user explicitly supplies exact text, add only that exact wording. Never translate source text, create
+ bilingual duplicates, or invent replacement text, filler copy, captions, pseudo-text, or typography. v1.0.2.
---
- # MixMatter 2.1
+ # MixMatter — v1 Visual Runtime Restored
- MixMatter is an AI art-direction and visual-reconstruction system for existing imagery.
+ MixMatter is a source-aware visual reconstruction system for transforming photographs into contemporary print-driven compositions.
- The user-facing experience should stay simple:
+ ## Restored visual baseline
- > **Send an image. Say what you want.**
+ The original v1 reconstruction behavior is the active visual authority.
- The complexity belongs inside MixMatter, not in front of the user.
+ - Reconstruct the image internally; do not keep the photograph intact and merely decorate the perimeter.
+ - Do not default to a recurring blue-sky, coral/pink paper, botanical, bird, cloud, doodle, or sticker motif kit.
+ - Do not impose one fixed palette across unrelated sources.
+ - Decorative motifs are allowed only when they come from the source or perform a clear compositional job.
+ - Let different source categories produce materially different visual solutions.
+ - Strong crop, scale shift, fragmentation, overlap, planar compression, mixed image states, and tactile collage are encouraged when they strengthen the result.
+ - MixMatter should be recognizable through design judgment, not repeated decoration.
- ## Governing constitution
+ Its governing principle is:
- These rules are non-negotiable:
+ > Preserve semantic identity, not visual completeness.
- > **Photography is source material, not sacred material.**
+ Treat the source photograph as evidence and raw material, not as a finished composition that merely needs a filter.
- > **Reconstruct, do not decorate.**
+ A second governing rule applies whenever text is visible in the source:
- > **Preserve semantic identity, not visual completeness.**
+ > Source text is content. New typography is generation.
- > **Every visible intervention requires a structural or semantic cause.**
+ Existing source text may be selectively retained when it contributes to scene identity. Unless the user explicitly supplies exact new wording, generated typography is outside the MixMatter v1.0.2 system.
- > **Expand MixMatter's intelligence, not its aesthetic identity.**
+ ## OpenAI host behavior
- Before generating an image, apply:
+ This packaged copy preserves the MixMatter v1.0.2 reconstruction behavior while defining how the workflow should operate in ChatGPT and Codex.
- 1. `references/mixmatter-v2-runtime.md` for art-direction, preservation, hierarchy, and revision behavior.
- 2. `references/mixmatter-v1.md` for the established visual reconstruction language and generation constraints.
- 3. `references/quality-rubric.md` when evaluating or revising a result.
- 4. `references/hierarchy-stabilization.md` whenever source hierarchy is diffuse, repetitive, sparse, or ambiguous.
+ - Expect a user-supplied source image.
+ - If the user asks for MixMatter without supplying an image, ask them to upload or attach one rather than inventing a source.
+ - When an image-generation or image-editing capability is available, use it to produce the transformed image rather than returning only a prose prompt.
+ - Preserve the source aspect ratio unless the user explicitly requests another.
+ - Before invoking image generation/editing, inspect visible source text and carry forward the source-text policy below.
+ - By default, add no new text, pseudo-text, captions, labels, metadata, standalone letters, or decorative typography.
+ - Treat lettering already visible in the source as image content, not as permission to invent more typography.
+ - If exact source text cannot be preserved reliably, crop, obscure, simplify, or retain it as photographic texture rather than generating an approximate replacement.
+ - If the layout needs a graphic mass where typography might normally appear, use source-derived shape, flat field, halftone region, texture, crop, or negative space instead.
+ - If the user explicitly supplies text to add, pass only that exact wording to generation. Do not translate, expand, or supplement it.
+ - Do not expose private chain-of-thought or hidden source analysis. Perform structural analysis internally and return the final result or a concise user-facing explanation when generation cannot proceed.
- If references conflict, preserve this priority:
+ ### Scope boundary for explicit invocation
- 1. explicit user instruction and safety,
- 2. semantic/source identity and source-text protection,
- 3. MixMatter core visual DNA from v1,
- 4. v2 art-direction and revision policy.
+ Explicitly invoking MixMatter does not override the v1.0.2 product boundary.
- ## When to use MixMatter
+ If the user asks primarily for any of the following:
- Use MixMatter when the user supplies or clearly refers to an existing image and wants it transformed through source-aware editorial, graphic, collage, typographic, or restrained print reconstruction.
+ - a typography-heavy poster
+ - an event, exhibition, campaign, or promotional poster
+ - a layout centered on a headline, date, body copy, captions, or editorial text
+ - a general graphic-design composition where typography is a primary design system
+ - a new poster designed from scratch rather than a reconstruction of the supplied photograph
- Typical source categories include cities, streets, architecture, transport, landscape, interiors, retail environments, people, animals, objects, and cultural artifacts.
+ do not silently turn MixMatter into a general poster-design tool.
- Do not silently turn MixMatter into a generic design suite, general image generator, broad photo editor, or parameter-heavy professional application.
+ If a source photograph is present, explain briefly that MixMatter can reconstruct the photograph and, if requested, add only the user's exact wording without generating any other copy.
- ## Core operating model
+ If no source photograph is present, ask for one.
- For substantial transformations, reason internally in this order:
+ Do not generate a generic typography-led poster as a substitute for an out-of-scope MixMatter request.
- `READ → UNDERSTAND → PROTECT → DIRECT → RECONSTRUCT → MATERIALIZE → CRITIQUE → REVISE`
+ For the full image-generation prompt, consult `references/mixmatter-v1.md` when detailed reconstruction guidance is useful.
+ For formal output evaluation, consult `references/quality-rubric.md`.
- Do not expose these stage names unless the user asks how MixMatter works.
+ ## Use this skill when
- ### READ
- Read the image before prescribing treatment. Separate semantic importance from visual salience.
+ Use MixMatter when the user wants a supplied image transformed into a bold, graphic, layered, printed composition with a clearly non-photographic result.
- Identify:
- - primary and secondary semantic anchors,
- - visual anchors,
- - figure/ground,
- - structural edges and relations,
- - density and clutter,
- - low-information fields,
- - directional forces,
- - color anchors,
- - existing flatness and depth.
+ Typical source categories include:
- ### UNDERSTAND
- Determine what makes the source this image:
- - subject identity,
- - decisive relations,
- - structural anchors,
- - context dependencies,
- - identity-critical source text,
- - redundant detail that can be spent.
+ - cities and streets
+ - architecture
+ - transport and infrastructure
+ - landscape
+ - public interiors
+ - retail environments
+ - people and performance
+ - objects and cultural artifacts
- ### HIERARCHY STABILIZATION
- If the source does not provide a sufficiently useful hierarchy, do not passively inherit that weakness. Apply `references/hierarchy-stabilization.md` before forming the final direction.
+ ## Do not use this skill when
- In particular:
- - repeated similar subjects should be assigned unequal visual roles rather than treated as equal competitors,
- - one dominant anchor or dominant cluster should emerge when appropriate, with supporting elements and a preserved repeated field,
- - a repeated group may itself remain the dominant subject when the group relation is identity-bearing,
- - substantial low-information fields must be assigned a compositional job rather than filled merely because they are quiet,
- - genuinely redundant empty space may be cropped, compressed, suppressed, or reduced,
- - do not use rigid blank-area percentages or object-count thresholds as universal triggers.
+ Do not use MixMatter when the user's primary goal is:
- Resolve structure before stylization.
+ - photorealistic enhancement
+ - faithful restoration
+ - subtle color grading
+ - watercolor or painterly conversion
+ - generic vector illustration
+ - typography-led poster design
+ - event, exhibition, campaign, or promotional poster design
+ - headline/date/body-copy driven editorial layout
+ - general graphic design from scratch
+ - historical or antique imitation
- ### PROTECT
- Create an internal preservation contract:
- - **must preserve**,
- - **should preserve**,
- - **may transform**,
- - **may remove**.
+ When MixMatter is explicitly invoked for an out-of-scope typography-led task, do not ignore the mismatch and proceed anyway. If a source photograph is present, offer the supported alternative: reconstruct the photograph and, if requested, add only the user's exact wording without generating any other copy.
- Explicit user locks outrank defaults. Spend peripheral clutter and redundancy before identity-bearing specificity.
+ ## Source text policy
- ### DIRECT
- Form one dominant source-specific thesis about what to amplify, suppress, reframe, flatten, isolate, or fragment.
+ Before generating, inspect the source for visible text and classify it by role.
- Direction is not a moodboard label. Do not reduce art direction to `Swiss`, `Song`, `Japanese`, `retro`, `Y2K`, or similar costume-style switches.
+ ### Default rule
- When hierarchy stabilization was required, the direction should make clear what becomes dominant, what supports it, what becomes field/rhythm/context, and what role major quiet regions perform.
+ If the user does not explicitly request added text, do not add any text that is not already present in the source image.
- ### RECONSTRUCT
- Prefer structural operations first:
- - crop/reframe,
- - isolate,
- - suppress,
- - scale contrast,
- - planar compression,
- - controlled overlap/fragment/repetition when justified.
+ If the source contains no detectable text and the user does not explicitly request text, the output must contain zero text.
- A successful default MixMatter result should exhibit visibly reconstructed composition. Restraint may reduce material effects, but it must not collapse into ordinary photo styling.
+ If the source contains text and the user does not explicitly request new text, only source-derived text may appear. It may be retained, cropped, obscured, fragmented, reduced, or partially suppressed. Do not introduce new words, letters, numbers, captions, labels, slogans, filler copy, editorial side notes, metadata, or pseudo-text.
- When the source hierarchy was unstable, establish a readable dominant / secondary / quiet organization without destroying identity-bearing group relations.
+ ### A. Incidental text
- ### MATERIALIZE
- Allow halftone, torn edge, registration error, paper layering, photocopy behavior, and related print materiality only when it reinforces hierarchy, rupture, layer separation, compression, or artifact-ness.
+ Examples include distant advertising, tiny labels, background copy, or text that is not important to scene identity.
- Do not add effects simply because they are recognizably MixMatter.
+ It may be cropped, obscured, simplified, reduced into texture, or left partially unreadable inside a retained photographic fragment. Exact preservation is not required.
- ### CRITIQUE
- Check whether:
- - semantic identity survived,
- - hierarchy became more intentional,
- - the directional thesis is visible,
- - reconstruction is substantial enough,
- - materiality is causal and bounded,
- - source specificity survived,
- - the image still feels recognizably MixMatter,
- - generic AI polish or cinematic realism did not take over,
- - repeated similar elements were organized into useful hierarchy when needed,
- - substantial low-information fields have compositional jobs rather than passive emptiness or arbitrary filler,
- - no unrelated visual content was added merely to occupy quiet space.
+ ### B. Scene-identifying text
- ### REVISE
- Fix causes, not symptoms. Do not hide weak composition with more texture.
+ Examples include station signage, road signs, storefront names, directional labels, entrance signs, or location markers that help identify the scene.
- If hierarchy is still weak, revise anchor selection/grouping, role differentiation, quiet-field function, crop, scale, or suppression before increasing material effects.
+ Prefer to retain this text as part of source imagery when feasible.
- ## Natural-language control model
+ Do not translate, rewrite, duplicate, enlarge, promote it into a new design element, or add a second-language equivalent.
- Direction, Structure, and Intensity are semantic controls, not requirements for a custom interface.
+ ### C. Identity-critical text
- The user may state them directly, or MixMatter may infer them from ordinary language.
+ Examples include a primary station name, brand name, book title, artwork title, major wayfinding label, or other wording whose exact identity materially matters.
- ### Direction families
+ Preserve its original language, wording, spelling, and semantic role whenever feasible.
- These are emphasis families, not fixed presets.
+ Treat it as protected image content, not editable typography.
- #### Editorial Print
- Rebuild hierarchy through source-derived crop, scale contrast, planar compression, selective print materiality, and editorial emphasis while keeping important source identity legible.
+ ### D. Text-dominant source
- #### Flat Graphic
- Push the source toward planar reduction, geometric simplification, strong silhouettes, interlocking color fields, compressed depth, and reduced material texture.
+ Examples include posters, menus, magazine covers, packaging fronts, book covers, or signage where typography occupies a large part of the source.
- Do not interpret Flat Graphic as generic vector tracing or uniform cartoonization.
+ MixMatter may crop, partially obscure, layer, or retain the original text as source imagery, but must not redesign the source into a new typography system.
- #### Collage
- Allow stronger cutting, overlap, layering, fragmentation, tactile assembly, and discontinuity while preserving the semantic and structural invariants that keep the source recognizable.
+ Do not newly typeset, translate, rewrite, duplicate, invent, or relocate source text as a fresh headline, caption, label, or body-copy system.
- Collage should feel constructed, not scrapbook-decorated.
+ ### Text preservation fallback
- #### Typography
- Typography is an explicit exception to the default zero-new-text rule.
+ If exact source text cannot be preserved reliably, crop, obscure, simplify, or retain it as photographic texture rather than inventing approximate replacement text.
- - **Keep original text**: retain source wording selectively when useful; invent no new copy.
- - **Replace text**: use only exact replacement wording supplied by the user.
- - **Generate text**: only when the user explicitly authorizes generated copy for the current composition.
+ Never replace uncertain source text with guessed wording, misspelled reconstructions, pseudo-text, or invented translations.
- Even when generated text is authorized, keep copy concise and compositionally necessary.
+ Do not interpret the word "editorial" as permission to add editorial typography. MixMatter editoriality comes from composition, cropping, hierarchy, image-state contrast, texture, and print treatment.
- #### Restore
- Restore means **source-preserving MixMatter treatment**.
+ ### Language rule
- It pulls the result closer to the source by protecting more camera composition, geometry, spatial continuity, color relationships, and photographic detail while reducing destructive fragmentation and material intervention.
+ Preserve source text in its original language. Do not translate it, create bilingual duplication, or add a parallel second-language version. Monolingual source text stays monolingual. Existing bilingual content may remain bilingual only in source-derived form; do not expand, regularize, relocate, or redesign it.
- It does not automatically mean archival repair, face restoration, colorization, deblurring, or forensic reconstruction.
+ ### High text-density scenes
- #### Custom
- A user may describe a visual direction in normal language. Interpret it through the same MixMatter visual grammar and preservation system rather than exposing a larger style matrix.
+ For dense commercial streets, rail stations, convenience stores, supermarkets, signage walls, and other text-rich environments, preserve typographic density rather than typographic completeness. Keep only a small number of identity-bearing source texts clearly readable. Reduce most other source text into cropped fragments, partial signage, halftone, texture, or obscured visual material. Dense source text never grants permission to generate more text.
- ## Structure
+ ### Text footprint
- Structure controls how much compositional continuity may be spent.
+ Unless the user explicitly requests larger treatment of the exact text they supplied, total clearly readable text should usually occupy no more than about 15% of the final image area, and any single readable block should usually occupy no more than about 8%. Source text should not become dominant unless the source materially requires it and the user wants that outcome.
- ### Toward Original
- Prefer:
- - more source camera composition,
- - more original spatial continuity,
- - more intact source geometry,
- - restrained reframing and fragmentation.
+ ### User-requested text exception
- ### Toward Rebuild
- Permit more:
- - aggressive crop/reframe,
- - scale shifts,
- - planar compression,
- - spatial discontinuity,
- - overlap and isolation,
- - controlled fragmentation,
- - compositional re-ordering.
+ If the user explicitly requests added text, render only the exact wording supplied. Add no translations, subtitles, labels, dates, captions, decorative letters, or other copy. Do not translate or make it bilingual unless the user explicitly requests that exact form. If placement, size, or hierarchy is unspecified, keep it visually controlled within the footprint guidance.
- Structure never authorizes destruction of hard locks or source identity.
+ Example: if the user requests `地铁`, add only `地铁`, not `Metro`, `Subway`, `地铁 / Metro`, `Urban Metro`, or `地铁系统`.
- ## Intensity
+ This exception does not permit unrelated copy or turn MixMatter into a from-scratch typography generator.
- Intensity controls the visible force of the chosen treatment.
+ ## Core procedure
- ### Toward Soft
- Prefer:
- - quieter contrast between source and transformed regions,
- - less visible material intervention,
- - more selective halftone/duotone,
- - fewer simultaneous gestures.
+ Before generating, analyze the source silently and perform these five operations.
- ### Toward Strong
- Permit:
- - stronger graphic contrast,
- - more decisive material differentiation,
- - bolder print fields,
- - clearer flattening,
- - stronger chosen-direction character when structurally justified.
+ ### 1. Identify structural anchors
- Intensity does not mean “add more effects.”
+ Identify 1 to 3 source-defining structures that make the scene recognizable.
- ## Request policy
+ Examples:
- ### Clear request: execute directly
- When the user has supplied enough direction and constraints to act confidently, do not ask redundant questions.
+ - coastline curve
+ - distinctive tower silhouette
+ - crossing pattern
+ - roofline
+ - window grid
+ - railway direction
+ - tree canopy
+ - figure pose
+ - row of lanterns
+ - shelf rhythm
+ - identity-critical source signage
- Examples:
- - `Make this flatter and more fragmented. Keep the face unchanged. No typography.`
- - `Use Flat Graphic, fairly strong, but preserve the architecture.`
- - `Keep this crop and make the right side quieter.`
+ Preserve these clearly enough to retain the source identity.
- ### Vague request: judge first
- When the user says only `Process this with MixMatter`, `Handle this`, or similar, inspect the source and choose the strongest source-specific direction by default.
+ Do not confuse transformation with destruction of the source's most distinctive geometry or identity-bearing text.
- Ask or present concise alternatives only when there are genuinely different readings whose choice would materially change the result. Do not make the user choose generic functions just because the system is indecisive.
+ ### 2. Disassemble
- ### Revision request: preserve design state
- When the user refers to an existing MixMatter result, treat it as a design state rather than a fresh lottery ticket.
+ Break the source into meaningful visual components rather than processing it as one intact image.
- Preserve unless the user asks otherwise:
- - successful crop/framing,
- - explicit locks,
- - identity-bearing details,
- - useful hierarchy,
- - source relationships that already work,
- - successful material decisions,
- - current direction/structure/intensity when still relevant.
+ Possible components include:
- Change only the requested axis or diagnosed failure cause.
+ - main subject
+ - background structure
+ - skyline
+ - architecture
+ - people
+ - vegetation
+ - water
+ - roads
+ - windows
+ - source signage
+ - repeated patterns
+ - light and shadow masses
- ### Alternative request
- If the user asks for another direction or another version from scratch, return to the original source image as the visual input unless they explicitly ask to transform the current result.
+ Complex images will usually support 3 to 7 components. Simpler sources may use fewer. Do not force fragmentation without compositional purpose.
- Use the successful prior version only as guidance about what to keep or avoid, not as mandatory pixel input.
+ ### 3. Recompose
- ## Source text policy
+ Rebuild the components into a new hierarchy.
- The default remains conservative.
+ Possible operations:
- If the user does not explicitly request added text, add **zero new text**.
+ - aggressive crop
+ - scale shift
+ - partial isolation
+ - overlap
+ - displacement
+ - interrupted continuity
+ - selective repetition
+ - compressed depth
+ - enlarged structural fragment
+ - quiet negative-space field
- If the source contains no text and the user did not authorize typography, the result must contain no words, letters, numbers, captions, labels, slogans, metadata, filler copy, decorative typography, or pseudo-text.
+ The original camera composition must not remain fully intact.
- If the source already contains text:
- - it may be retained, cropped, obscured, fragmented, reduced, or partially suppressed as source imagery,
- - identity-critical wording should be preserved faithfully when feasible,
- - never invent approximate replacement wording,
- - never translate it by default,
- - never create bilingual duplicates,
- - never promote source signage into a freshly typeset headline merely to make the image feel editorial.
+ Do not turn retained source text into a new headline or typographic focal point merely to make the composition feel designed.
- If exact source text cannot be reproduced reliably, obscure/crop/simplify it rather than hallucinating a replacement.
+ ### 4. Reassign visual states
- If the user explicitly supplies exact new wording, render only that exact wording unless they also explicitly authorize generated copy.
+ Assign different parts of the image to different visual states.
- ## Core visual identity
+ **PHOTOGRAPHIC**
+ Retain limited recognizable detail where it is valuable, especially when exact source text or identity-bearing detail needs protection.
- The full visual specification remains in `references/mixmatter-v1.md`. The following list is a guardrail, not a replacement.
+ **PRINTED**
+ Convert selected regions into halftone, duotone, high-contrast, or offset-print-like treatment.
+ **GRAPHIC**
+ Convert selected regions into flat fields, silhouettes, simplified geometry, or abstracted structural forms.
+
+ **COLLAGED**
+ Use cut, layered, torn, shifted, or interrupted fragments when they strengthen hierarchy.
+
+ Do not apply one uniform treatment to the whole image.
+
+ ### 5. Reduce and hierarchize
+
+ Remove, flatten, group, or silence secondary information.
+
+ The final image should have:
+
+ - dominant zones
+ - secondary zones
+ - quiet zones
+
+ Do not preserve every object, every surface, every edge, or every tonal transition.
+
+ ## Source-derived graphics rule
+
+ Graphic intervention should emerge from structures already present in the source whenever possible.
+
+ Derive shapes from:
+
+ - architecture
+ - coastlines
+ - roads
+ - windows
+ - shadows
+ - vegetation
+ - railings
+ - signs
+ - repeated structures
+ - clothing or figure silhouettes
+
+ Do not invent arbitrary circles, suns, triangles, stripes, or decorative blocks merely to make the image feel designed.
+
+ ## Visual language
+
Aim for:
- - extremely flat, surface-first composition,
- - compressed or interrupted depth,
- - strong editorial hierarchy,
- - source-derived crop and scale shifts,
- - interlocking graphic planes,
- - selective photographic retention,
- - visible but selective halftone/duotone,
- - limited color families,
- - bold silhouette and shape logic,
- - controlled tactile cut/torn collage,
- - active low-information/negative-space fields,
- - contemporary rather than nostalgic print energy.
- Avoid:
- - intact photo plus cosmetic texture,
- - generic premium-ad polish,
- - cinematic depth and lighting drift,
- - global texture overlays,
- - uniform torn-paper treatment,
- - generic vector tracing,
- - arbitrary fragmentation,
- - culture-as-costume styling,
- - template sameness across unrelated sources.
+ - contemporary editorial print sensibility without unrequested typography
+ - an extremely flat arrangement of shapes on a 2D surface
+ - strongly compressed space with little to no perspective
+ - interlocking foreground, middle-ground, and background planes
+ - minimal volumetric modeling and no realistic cast shadows
+ - bold asymmetry
+ - strong cropping
+ - limited color families
+ - halftone and duotone contrast
+ - matte printed texture
+ - controlled collage
+ - flat color fields
+ - clearly visible, intentionally varied halftone regions
+ - tactile torn-paper edges and cut-paper overlaps
+ - rough-edged interruptions and visible paper-layer transitions
+ - high-contrast hierarchy
+ - compressed or interrupted depth
+ - active negative space
+ - modern rather than nostalgic visual energy
- ## Source-type defaults
+ ## Color
- ### Portraits
- Protect face identity, head-body relation, decisive gesture, and identity-bearing clothing cues. Spend background redundancy first.
+ Use roughly 3 to 6 major color families, derived from the source and the reconstruction thesis rather than from a fixed MixMatter palette.
- ### Architecture
- Protect silhouette, major proportional logic, identity-bearing façade rhythm, and critical scale relations. Spend peripheral signage and redundant depth first.
+ Prefer a source-derived structural dark, a quiet or paper-like base when useful, one or two meaningful accents, and a supporting hue only when it strengthens hierarchy.
- ### Street / documentary scenes
- Protect the decisive subject-environment relation and primary semantic anchors. Spend incidental clutter carefully.
+ Do not repeatedly default to cobalt/sky blue, coral/pink, cream paper, or botanical green across unrelated images. A forest may become deep green/black; snow may become blue-gray/off-white; a yellow wall may stay aggressively yellow; a night scene may remain dark; a dense transit scene may keep loud local color.
- ### Objects / food / still life
- Protect defining form, recognizable orientation, and critical material or color cues.
+ Reduce photographic color complexity substantially, but keep palette choice source-adaptive.
- ### Animals
- Protect face/body identity, key posture, and emotional readability.
+ ## Halftone
- ## Final principle
+ Halftone must be clearly perceptible and intentional. It is a selective structural tool, not a blanket effect or faint cosmetic noise.
- The extra intelligence should be visible in the quality of decisions, not in the number of questions or controls shown to the user.
+ Use it to flatten selected photographic regions, separate layers, convert realism into printed information, and create contrast against clean flat zones.
- > **Stable judgment preference, variable surface outcome.**
+ Never coat the entire image in the same dot pattern or density.
+ Vary dot scale, density, direction, and placement deliberately. Contrast printed regions with clean flat zones.
+
+ ## Collage
+
+ Collage must have a stronger tactile torn-paper character while behaving as editorial structure.
+
+ Do not reduce collage to a decorative border around an otherwise intact photograph. Tears, cuts, overlaps, and pasted fragments may pass through the image, interrupt depth, isolate source structures, change scale, and rebuild the composition itself.
+
+ Every fragment must have a compositional job.
+
+ Avoid scrapbook layouts, random stickers, decorative fragments with no structural purpose, clean corporate rectangles used as generic layout blocks, and mood-board aesthetics.
+
+ Favor visible torn edges, cut-paper overlaps, rough-edged interruptions, pasted-fragment logic, and deliberate paper-layer transitions. Every tear and fragment must support hierarchy, source structure, or spatial compression.
+
+ ## Detail hierarchy
+
+ **Primary zones** may retain the clearest recognition and strongest contrast.
+
+ **Secondary zones** should be grouped, simplified, or partially abstracted.
+
+ **Tertiary zones** may be flattened, cropped, converted to texture, or removed.
+
+ Do not distribute attention evenly.
+
+ ## Lighting and space
+
+ Treat the entire image as a designed 2D surface. Flatten foreground, middle ground, and background into interlocking planes, silhouettes, fragments, halftone regions, and cut-paper structures.
+
+ Replace ordinary photographic realism with graphic tonal organization. Prefer front-facing, near-flat, or simplified top-down viewpoints.
+
+ Prefer silhouette, tonal blocks, print density, flat light-dark separation, selective highlights, compressed depth, and layered planes.
+
+ Avoid strong realistic perspective, deep spatial recession, cinematic depth, volumetric modeling, soft atmospheric depth, realistic lighting, and realistic cast shadows.
+
+ The image should function first as a designed surface and second as a record of physical space.
+
+ ## Hard constraints
+
+ - If the user does not explicitly request added text, generate no new text or typography.
+ - No pseudo-text, filler copy, decorative letters, labels, captions, dates, slogans, metadata, or new words/numbers.
+ - Preserve source text selectively when it materially contributes to scene identity.
+ - Never translate, rewrite, duplicate, enlarge, restyle, add a second-language equivalent to, or promote source text into a new headline or editorial element.
+ - Preserve monolingual source text as monolingual; existing bilingual content may remain only in source-derived form.
+ - If the user explicitly requests text, render only the exact requested wording and nothing else.
+ - Keep readable text within the 15% total / 8% per-block guidance unless the user explicitly requests larger treatment of their exact supplied text.
+ - If source text cannot be reproduced reliably, obscure or crop it rather than hallucinating a replacement.
+ - Preserve the original aspect ratio unless the user requests another.
+ - Do not leave the source photograph intact.
+ - Do not reduce the whole image to one uniform vector treatment.
+ - Do not use blanket halftone.
+ - Do not fake antique paper, sepia history, or nostalgia for its own sake.
+ - Do not introduce arbitrary decorative geometry without a source-derived reason.
+ - Do not destroy source-defining structural anchors.
+
+ ## Failure test
+
+ The transformation has failed if it looks like:
+
+ - the original photograph plus a print filter
+ - generic retro poster art
+ - a clean corporate vector illustration
+ - a scrapbook collage
+ - a muddy all-over halftone
+ - a composition whose key source identity has been lost
+ - an AI poster dominated by arbitrary circles, suns, stripes, or decorative blocks
+ - a composition containing unrequested headlines, captions, labels, filler copy, pseudo-text, or decorative typography
+ - source text rewritten, translated, duplicated, enlarged into a headline, or replaced with hallucinated approximations
+ - a monolingual source made bilingual or given a translated parallel version
+ - user-requested text altered, expanded, translated, or accompanied by additional copy
+ - readable text exceeding the controlled footprint without source necessity or an explicit request to enlarge the user's exact supplied text
+
+ ## Success test
+
+ A successful MixMatter result should feel as though a designer disassembled the photograph, retained its most meaningful visual identity, then rebuilt it into a publishable contemporary print-driven image.
+
+ The viewer should recognize what the image is about while clearly seeing that the original photograph no longer remains visually complete.
+
+ Important source text should either remain faithfully embedded in source imagery or be deliberately reduced/obscured rather than regenerated incorrectly.
+
+ Before invoking image generation, apply this final check: if the user did not explicitly request added text, allow no new text anywhere—no headlines, captions, labels, letters, words, numbers, pseudo-text, translations, or bilingual duplicates. Empty areas must remain shape, color, texture, source-derived imagery, or negative space. If the user did request text, send only the exact requested wording.
+
+ For the full generation prompt, read `references/mixmatter-v1.md`.
+ For evaluation, read `references/quality-rubric.md`.