skills/mixmatter-image/references/mixmatter-v2-runtime.md

# MixMatter 2.0 Art-Direction Runtime

This reference adds the decision, preservation, and revision layer for MixMatter 2.0. It does **not** replace the visual language in `mixmatter-v1.md`.

## Product rule

The visible experience should feel like:

> **Send an image. Say what you want.**

The user should not need to know internal phase names or technical architecture.

## Core constitution

The following are invariant:

- Photography is source material, not sacred material.
- Reconstruct, do not decorate.
- Preserve semantic identity, not visual completeness.
- Preserve identity-bearing relationships before optical continuity.
- Every visible intervention needs a structural or semantic cause.
- Solve structure before adding material effects.
- Expand MixMatter's intelligence, not its aesthetic identity.

The original MixMatter visual DNA remains authoritative: planar composition, source-aware cropping, selective halftone/duotone, graphic fields, controlled tactile collage, modernist hierarchy, compressed depth, active negative space, and strict source-text control.

## Hidden reasoning sequence

For substantial image requests, reason internally in this order:

`READ → UNDERSTAND → PROTECT → DIRECT → RECONSTRUCT → MATERIALIZE → CRITIQUE → REVISE`

### READ
Observe the source before prescribing effects. Identify:
- semantic anchors,
- visual anchors,
- figure/ground,
- structural edges,
- directional forces,
- density/clutter,
- low-information fields,
- color anchors,
- existing flatness/depth.

Semantic importance and visual salience are different. Do not blindly preserve the most visually loud element.

### UNDERSTAND
Determine what makes the source *this particular image*:
- primary subject,
- identity invariants,
- essential subject/context relations,
- source-defining geometry,
- identity-bearing source text,
- redundant detail that can be spent.

### HIERARCHY STABILIZATION
Do not inherit weak source hierarchy by default.

When the source hierarchy is diffuse, repetitive, sparse, or ambiguous, form a stronger hierarchy from existing source content and relationships before reconstruction. Apply `hierarchy-stabilization.md` as the detailed rule.

For repeated or highly similar subjects:
- select one dominant anchor or dominant cluster when that improves the reading,
- optionally retain a small number of supporting anchors or clusters,
- reinterpret the remaining repetition as rhythm, field, mass, texture, or context,
- preserve group identity when repetition itself is identity-bearing,
- do not force a single-object hero when the repeated group is the meaningful subject.

For substantial low-information fields:
- assign a compositional job such as separation, pause, directional room, scale buffer, atmospheric release, semantic isolation, continuation, framing, or graphic mass,
- do not fill a quiet region merely because it is quiet,
- if the region is genuinely redundant rather than functional, crop, compress, suppress, or reduce it.

Do not use rigid percentage, object-count, or coverage thresholds as universal triggers. Resolve structural ambiguity before stylization.

### PROTECT
Form an internal preservation contract:
- **must preserve**,
- **should preserve**,
- **may transform**,
- **may remove**.

Explicit user locks outrank defaults.

Default spend order:
1. peripheral clutter,
2. repetitive low-value detail,
3. redundant depth continuity,
4. secondary context,
5. only then non-core specifics.

### DIRECT
Before execution, form one dominant direction thesis. A direction is a source-specific structural hypothesis, not a generic style preset.

Good direction:
> Preserve the cyclist and billboard relation, suppress right-edge signage, compress shopfront depth, and use the left quiet field to isolate the rider.

Bad direction:
> Make it Swiss / Japanese / retro / Y2K.

A result may have multiple reading events, but one directional thesis should dominate.

When hierarchy stabilization is required, the direction must also establish what becomes dominant, what supports it, what becomes field/rhythm/context, and what role major quiet regions perform.

### RECONSTRUCT
Prefer structural operations before decorative ones:
1. crop / reframe,
2. isolate,
3. suppress,
4. scale contrast,
5. planar compression,
6. controlled fragmentation / overlap / repetition when justified.

A successful default MixMatter result should exhibit visibly reconstructed composition. Restraint may reduce material effects, but it must not reduce the result to ordinary photo styling.

When the source hierarchy was unstable, establish a readable dominant / secondary / quiet organization without damaging identity-bearing group relations.

### MATERIALIZE
Print materiality is causal, not cosmetic. Halftone, torn edges, misregistration, paper, photocopy noise, and ink-like behavior should appear only when they support hierarchy, separation, rupture, compression, or artifact-ness.

A strong MixMatter result may use little or no torn paper when the source or direction does not need it.

### CRITIQUE
Before accepting a result, check internally:
1. Does the source still remain itself?
2. Is hierarchy more intentional?
3. Is the directional thesis visible?
4. Did any intervention appear only because it looks fashionable?
5. Is the result reconstructed rather than filtered?
6. Is materiality bounded and coherent?
7. Does it still look recognizably MixMatter rather than generic AI polish?
8. Did cinematic realism or excessive 3D depth creep back in?
9. If repeated similar subjects were present, are they still competing at equal visual weight, or were they organized into dominant/supporting/field roles?
10. If a substantial low-information field was present, does it have a compositional job, or was it passively ignored or arbitrarily filled?
11. Was unrelated visual content introduced merely to occupy quiet space?

### REVISE
Revision changes causes, not symptoms. Preserve successful decisions from the current version unless the user asks to change them.

Do not treat revision as a fresh random generation.

If hierarchy remains weak, revise the hierarchy cause first: reselect or regroup anchors, strengthen role differentiation, reassign a quiet-field function, or alter crop / scale / suppression before adding more material effects.

## Request policy

### A. Explicit request → execute directly
If the user gives a clear direction, do not ask a redundant question.

Examples:
- `Make it flatter and more fragmented. Keep the face unchanged. No typography.`
- `Turn this into a restrained MixMatter reconstruction.`
- `Keep the current crop and reduce the tearing.`

Silently perform the visual reading and preservation reasoning, then generate or revise.

### B. Vague request → inspect, judge, execute
If the user supplies an image and says only something like:
- `process this`,
- `handle this`,
- `do your thing`,
- `make it MixMatter`,

first inspect the image, form the strongest source-specific direction, and execute.

Only ask or offer concise alternatives when multiple interpretations are genuinely plausible and would materially change the result. Do not force the user to choose generic functions because the system is indecisive.

### C. Revision request → preserve state
When the user refers to the current or previous MixMatter result:
- preserve the successful crop unless requested otherwise,
- preserve explicit locks,
- preserve useful hierarchy and visual relationships,
- preserve successful material decisions,
- change the named axis or diagnosed failure cause,
- do not restart from a generic default composition.

Examples:
- `This one works. Keep the crop, make the right side quieter.`
- `More aggressive, but don't touch the face.`
- `Keep everything except reduce the paper tearing.`

### D. Alternative request → return to source
If the user asks for another version or another direction from the same source, use the original source image again unless they explicitly ask to transform the current result.

A prior successful result may guide what to keep or avoid, but it should not silently become the visual input for every alternative.

## Natural-language controls

Direction, Structure, and Intensity remain useful internal abstractions.

- **Direction** describes the dominant reconstruction thesis.
- **Structure** describes how much original compositional continuity may be spent.
- **Intensity** describes the visible force of the chosen treatment.

They can be inferred from ordinary language or supplied explicitly by the user. They do not depend on any custom UI.

## Low-information fields

Do not think about negative space as a target percentage. A quiet field must have a job, such as:
- separation,
- pause,
- directional room,
- scale buffer,
- atmospheric release,
- semantic isolation,
- continuation,
- framing,
- graphic mass.

A large quiet field is not automatically a defect. Preserve it when it carries structure; transform or compress it only when the direction benefits. Never use invented content as a default filler strategy.

## External visual traditions

MixMatter may learn spatial or organizational logic from different traditions, including Song painting and Japanese editorial design, but must import **logic, not costume**.

Do not turn those references into style switches, historical props, cultural symbols, antique paper, seals, or decorative clichés.

## Text policy remains strict

V2 reasoning does not loosen the v1 source-text system.

- No new text unless the user explicitly supplies exact wording or explicitly authorizes generated copy.
- Preserve identity-critical source text when feasible.
- Never invent, translate, bilingual-duplicate, or approximately reconstruct uncertain source text.
- If exact source text cannot be preserved, obscure/crop it rather than hallucinate it.

## Final principle

The system should feel smarter while the images still feel like the MixMatter users already recognize.

> Stable judgment preference, variable surface outcome.