Imported from mihirduvedi/master-design-studio (skills/humanize-image-generation/SKILL.md). Install upstream with npx skills add mihirduvedi/master-design-studio --skill humanize-image-generation. Copyright stays with the author (MIT).

Humanize Image Generation

Treat the model as one production tool inside a directed visual process. A human-made feel comes from specific intent, coherent capture or craft logic, selection, and repair—not from adding random defects or merely saying “not AI.”

Use the client's raster image generation or editing capability and follow its input-image, output-path, transparency, provenance, and iteration rules. Load a dedicated image-generation skill when available. Prefer native SVG, vector, 3D, or layout tools when the requested asset needs exact editable geometry, typography, or a reusable system.

Resolve every bundled relative path from the directory containing this SKILL.md; never interpret references/... or scripts/... from the user's project directory. The bundled scripts/compare_edit_regions.py requires Python 3.11+ and Pillow; check availability before using it and do not claim the comparison ran when the dependency is unavailable.

Required workflow

1. Define truth status and provenance. Distinguish an invented image from documentary evidence. Preserve generation metadata and provenance signals. Never remove watermarks or provenance, imply that a synthetic event actually happened, fabricate evidence, or describe generated media as camera-original.
2. Resolve the job. Record the audience, purpose, final placement, crop, dimensions, subject, action, setting, emotional register, and non-negotiable content. Inspect supplied references before inventing a direction.
3. Research close visual precedents. If supplied references are insufficient and external research is allowed and available, use dedicated image search before prompting. Read reference-research.md, build advanced queries from subject, action, viewpoint, light, medium, material, era, and use, then inspect the original source pages. Shortlist a small role-labeled board rather than one style target. Follow a single reference closely only when the user owns it, supplied it for transformation, or its license/status permits that use; otherwise synthesize transferable traits across multiple references without reproducing distinctive protected expression.
4. Choose one production grammar. Select photography, illustration/cartoon, drawing/painting/print, 3D/product rendering, or designed element/composite. Read the matching section of medium-playbooks.md. Do not blend unrelated media signals by default.
5. Write a capture or craft contract. Define the human decisions that would plausibly produce the image before writing the model prompt. Map each approved reference to one role such as composition, gesture, light/color, object mechanics, or material/marks.
6. Lock approved masters. When the user identifies an approved image as the master, freeze its file, dimensions, composition, camera, geometry, and all non-target regions. Define the permitted edit region or semantic mask before editing, select a route capable of preserving it, and verify non-target pixels or structures afterward. Read authorship-and-preservation.md. A prompt that says “keep everything else unchanged” is not a preservation mechanism.
7. Reduce scene risk. Plan exact subject count, object relationships, hand-object interactions, foreground/background order, and required negative space. Split dense layouts, exact typography, repeated characters, or complex composites into controllable passes when one-shot generation is unreliable.
8. Prompt with observable causes. Describe who or what, doing what, where, viewed from where, under what motivated light, through which medium, and with which material or capture consequences. Label input images by role and state what must not be copied. Prefer concrete nouns, verbs, spatial relations, and surface behavior to stacks of quality adjectives.
9. Generate for selection. When time and tool budget allow, create a small set of genuinely different candidates, not cosmetic recolors. Curate for story, composition, and structural credibility before polishing texture. If only one candidate is practical, plan a deliberate edit pass.
10. Audit at three scales and by material region. Judge the silhouette and authorship at a glance, the composition and medium consistency at normal size, then high-risk details enlarged. For products and renders, compare body, controls, glass, metal, wood, fabric, and emitted light separately; unrelated materials must not share one generic polish treatment. Compare against the reference roles, not just aesthetic resemblance. Apply artifact-audit.md. If one deliberate refinement retains the same systemic artificial surface language, stop refining through the same generator and apply the route-switch gate in authorship-and-preservation.md.
11. Repair locally. Make one targeted change at a time. Re-state all invariants. Prefer a zoomed crop or localized edit for hands, faces, text, object mechanics, edges, reflections, and material response; do not regenerate the whole image after the composition is approved unless local repair cannot preserve it.
12. Finish in the right tool. Typeset long or exact copy in a layout/vector tool, fix masks and seams, align geometry, and apply restrained regional crop and color finishing. Do not use grain, blur, texture, or compression to conceal unresolved structural errors.
13. Verify the placed asset. Inspect the exact final file at its real crop and consumption size, not only the raw generation. For cutouts, inspect alpha corners and edges, trim transparent padding, preview against every intended background, and verify that the asset's apparent scale and shadow make sense in the composition. Recheck important details after resizing, masking, compositing, or export.
14. Deliver traceably. Return the selected asset, note the final prompt and material edits when useful, and keep a compact reference/provenance log with the project.

Capture or craft contract

Write these decisions in working notes:

Intent: <what the image communicates and why it exists>
Moment: <specific action or visual beat, not a generic pose>
Reference map: <source or Image 1/2/3 -> one permitted influence each>
Medium: <plausible capture or making process>
Composition: <viewpoint, crop, subject placement, occlusion, negative space>
Light and color: <motivated source, tonal range, restrained palette logic>
Surface and marks: <medium-specific texture or material response>
Reality rules: <anatomy, mechanics, perspective, physics, culture or period>
Controlled irregularity: <1-3 causal, local, story-serving imperfections>
Restraint: <what stays simple, quiet, soft, cropped, or absent>
Invariants: <facts and visual properties that must not drift>
Edit authorization: <approved master; exact region or property allowed to change>

If a choice cannot be explained as capture, craft, story, or function, remove it.

Prompt discipline

Use this order as scaffolding, omitting empty lines:

Use and truth status: <placement; fictional/conceptual/editorial>
Primary request: <subject + action + purpose>
Reference inputs: <Image 1 composition only; Image 2 light/color; Image 3 material/marks; exclusions>
Scene and relationships: <setting; exact spatial and object logic>
Medium/process: <photo, ink, gouache, cel animation, 3D, collage, etc.>
Composition: <viewpoint, crop, focal plane, negative space>
Light/color: <source direction and palette behavior>
Material/mark behavior: <where texture appears and why>
Controlled irregularity: <small causal details>
Constraints: <must preserve; exact count; factual requirements>
Avoid: <request-specific failure modes>

Avoid generic quality soup such as “masterpiece, stunning, ultra-detailed, 8K, HDR, cinematic” when it substitutes for decisions. Prompt length alone does not create quality. Use “cinematic,” “professional,” “handmade,” or “realistic” only when the contract explains what that means in visible terms.

Human irregularity standard

Use irregularity as evidence of process, not decoration:

• Put wear where hands, weather, friction, gravity, or time would create it.
• Let fabric crease at joints, pigment pool at wet edges, pencil pressure vary through a stroke, and roughness change where surfaces are touched.
• Preserve selective focus, edge hierarchy, overlap, occlusion, and partial information when the medium calls for them.
• Allow asymmetry when pose, manufacture, or composition produces it; preserve intentional symmetry when the subject requires it.
• Keep some regions quiet. Uniform micro-detail, universal sharpness, and evenly distributed distress look procedural.
• Never add malformed anatomy, broken mechanics, false text, random grime, or arbitrary jitter as “human imperfection.”

Hard failure conditions

Reject or repair an image when any of these remains:

• anatomy, pose, gaze, or hand-object contact is implausible;
• an object cannot function, attach, fasten, stand, or be used as depicted;
• light, shadow, reflection, perspective, scale, depth, or contact contradicts the scene;
• text, numerals, symbols, counts, or required spatial relationships are wrong;
• medium logic changes arbitrarily across the frame;
• an approved master changes outside the authorized region or loses its camera, geometry, identity, or established composition;
• unrelated materials collapse into one hue family, one texture scale, one roughness value, or one generic highlight treatment;
• global smoothing, uniform sharpening, blanket grain, or generic “premium” polish replaces meaningful regional variation;
• procedural “authenticity” texture—repeated speckles, pseudo-grain, evenly distributed scuffs, or uniform distress—has been added to disguise a systemic generative look;
• waxy skin, plastic hair, over-gloss, hyper-saturation, uniform detail, smudged boundaries, or local resolution mismatches dominate;
• cultural, historical, uniform, ritual, or contextual details are unsupported or stereotyped;
• the composition is generic, overly centered, decorative, or polished without a content reason;
• a living artist is imitated, a protected work or distinctive reference arrangement is reproduced, reference provenance is missing for a close transformation, or media provenance is misleading;
• the final placed crop has not been inspected.
• a cutout retains opaque corners, key-color fringe, excessive transparent padding, or a baked shadow that conflicts with its destination.

Reference map

• Read medium-playbooks.md for photography, illustration/cartoon, drawing/painting/print, 3D rendering, and designed-element workflows.
• Read reference-research.md before external image discovery, close-reference selection, reference-board construction, or reference-led prompting.
• Read artifact-audit.md before candidate selection, repair, and release.
• Read authorship-and-preservation.md before editing an approved master or diagnosing global polish, color collapse, and material sameness.
• Read research-foundations.md when evolving the skill, resolving conflicting guidance, or explaining why a rule exists.