calesthio/visual-style-direction
Translate a creative brief, brand system, or reference set into an original, portable visual-direction system for image and video production. Use for art direction, reference analysis without imitation, design tokens, palette and contrast, typography, composition, texture and material, lighting, lens and camera language, motion language, cross-shot continuity, accessibility, production handoff, visual review, and repair; do not use for provider-specific API syntax or model-specific prompt recipes.
npx skills add https://github.com/calesthio/generative-media-skills --skill visual-style-directionCreate a visual grammar that another production agent can execute repeatedly, not a pile of adjectives and not a clone of a reference. Translate intent into observable rules, encode the rules in portable artifacts, prove them on representative frames or shots, and repair the system rather than patching isolated outputs.
Keep these categories distinct in the direction packet. Never present taste, convention, or a provider behavior as a standard.
Produce a visual direction packet that is usable without access to the original conversation. It must contain:
10. representative proof-frame or proof-shot specifications;
11. handoff metadata, acceptance tests, and a repair log.
Do not bury decisions in prose. Every consequential choice needs a stable name, rationale, allowed range, and test.
Use this skill to decide what the work should look and move like across tools. Hand the resulting system to image-generation, video-generation, animation, editing, layout, or finishing skills.
Do not:
Extract the following before proposing a direction:
| Area | Questions to resolve |
|---|---|
| Communication | What must the viewer understand, feel, remember, or do? |
| Audience | Who is included, excluded, represented, or potentially misread? |
| Brand | Which supplied assets and rules are authoritative? Which are only examples? |
| Deliverables | Still, sequence, film, social cutdowns, thumbnails, UI, print; required ratios and durations? |
| Production | Live action, generated media, 2D/3D, compositing, edit, or a hybrid? What can actually be controlled downstream? |
| Continuity | Which subjects, products, spaces, props, wardrobe, marks, and facts recur? |
| Accessibility | Captions, audio description, reduced-motion version, text alternatives, contrast target, reading conditions? |
| Rights | Source, owner, license, territory, term, model/property releases, trademark constraints, synthetic-media disclosure? |
| Color | SDR/HDR, display/print targets, working space, output transform, embedded profile, review environment? |
| Approval | Who approves the route, proof frames, continuity sheet, and final master? |
If a missing answer changes the direction materially, ask. If it only changes an implementation detail, state a reversible assumption and continue.
Record every reference with:
Use descriptions such as “low-key side light with rapid falloff and a small warm practical in the background,” not “cinematic like Reference A.”
Transfer purpose and general mechanisms. Rebuild surface choices from the new brief. Exclude identity-bearing details unless supplied and licensed for use.
Research finding: Controlled design-fixation experiments found that examples can cause designers to reproduce example features and narrow conceptual output. The finding originated in engineering-design tasks, so use it as a warning about reference exposure, not proof that every mood board causes fixation. Counter it by analyzing multiple functionally different sources, delaying surface decisions until the brief is stated, and documenting exclusions. See Jansson and Smith, “Design Fixation,” *Design Studies* 12(1), 1991 (DOI page90003-F)).
Ask:
Documented fact, U.S. context: The U.S. Copyright Office notes that style standing alone is generally not protected by copyright, while original pictorial and graphic expression can be protected and permission may be needed to use another person's work. That distinction does not make imitation risk-free: individual expressive elements, derivative-work rights, contract, publicity, and trademark issues remain. The USPTO explains that confusing similarity can arise from appearance, meaning, or overall commercial impression. Use this as a risk screen, not legal advice. See the Copyright Office Digital Replicas report, Copyright for Visual and Graphic Artists, and USPTO likelihood-of-confusion guidance.
When a user names a living artist or asks for a near-copy, decline the identity-bearing imitation and offer a mechanism-based direction using broader era, medium, lighting, composition, material, and emotional attributes.
Propose two to four routes that solve the same communication problem through meaningfully different systems. Change at least three structural dimensions across routes, such as:
Do not make “Route B” a color swap of Route A.
For each route, provide:
Recommend one route against the brief. Preserve rejected routes as decision history; do not silently blend them into the approved route.
The thesis must contain a communication intent, a visual tension, and a behavioral rule.
Weak: “Premium, modern, cinematic.”
Strong: “Make technical precision feel tactile: rigid editorial geometry contains warm, imperfect materials, while the camera moves only when a relationship becomes clearer.”
Translate the thesis into:
| Element | Invariant | Controlled variable | Failure signal |
|---|---|---|---|
| Subject | recognizable silhouette, product geometry | pose, crop, expression | anatomy/mark changes between shots |
| Composition | one dominant mass, upper-left text reserve | subject scale 35–60% frame height | equal-weight clutter |
| Light | large cool key, narrow warm practical | key angle ±15°, practical position | flat frontal fill or mixed-white imbalance |
| Motion | purposeful translation, no idle drift | amplitude by beat importance | constant motion with no information role |
Use tokens for reusable atomic and semantic values. Use prose or structured ledgers for relationships that are not reducible to a value, such as “camera movement must reveal causality.”
Separate:
color.text.primary, color.surface.hero, motion.duration.emphasis;color.text.on-dark only over surfaces passing the target contrast test.”Avoid names tied to a current value (blue-500) when the role matters (accent-action). Keep source values and output-space values distinct.
When a color token carries redundant representations such as numeric components and a fallback hex, validate that they encode the same color in the declared color space. Prefer one authoritative representation over contradictory convenience fields. Run a token parser or color-conversion check before handoff; a syntactically valid token with mismatched values is not portable.
Documented fact: The Design Tokens Community Group describes tokens as a platform-agnostic way to exchange design decisions. Its first stable format, 2025.10, uses typed tokens with $value, supports groups and references, and defines types including color, typography, dimension, duration, stroke, gradient, and shadow. Verified 2026-07-09. Follow the stable specification when interchange is required; keep camera, lighting, composition, and continuity ledgers beside it rather than inventing nonstandard token types. See DTCG Format Module 2025.10.
This is an example, not a mandatory palette.
{
"color": {
"ink": {
"$type": "color",
"deep": {
"$value": {
"colorSpace": "srgb",
"components": [0.055, 0.071, 0.09],
"hex": "#0E1217"
}
}
},
"text": {
"$type": "color",
"on-dark": {
"$value": {
"colorSpace": "srgb",
"components": [0.957, 0.941, 0.902],
"hex": "#F4F0E6"
}
}
}
},
"motion": {
"duration": {
"$type": "duration",
"settle": { "$value": { "value": 420, "unit": "ms" } }
}
},
"type": {
"title": {
"$type": "typography",
"$value": {
"fontFamily": ["Approved Display Family", "sans-serif"],
"fontSize": { "value": 64, "unit": "px" },
"fontWeight": 650,
"lineHeight": 1.02
}
}
}
}
Also provide a human-readable token table with role, value, allowed use, prohibited use, contrast partners, and fallback.
Label numeric layout, type, motion, and safe-area ranges as proof-set starting points until they have been tested against the actual regulated copy, device sizes, platform UI, viewing distance, and encoded deliverables. Promote them to approved constraints only after the relevant proof gate passes.
Define color by function and relationship:
Specify dominant/support/accent proportions as flexible ranges, not decorative arithmetic. Test all important pairs on actual textured, graded, and moving backgrounds—not only flat swatches.
Documented fact: WCAG 2.2 Level AA specifies at least 4.5:1 contrast for normal text, 3:1 for large text, and 3:1 for graphical objects needed to understand content. It also states that color must not be the only visual means of conveying information. These are web-content criteria, not a universal cinema-grading standard; use them as minimum acceptance tests for web/UI overlays, captions, diagrams, and related assets, then test in the real delivery environment. See WCAG 2.2 §§1.4.1, 1.4.3, 1.4.11.
For moving footage, evaluate worst-case frames. Add a backing plate, local darkening, outline, shadow, repositioning, or alternate color token when a pair fails. Never “solve” contrast solely by making everything white or black if hierarchy is lost.
Record:
Documented fact: ICC profiles connect color encodings across devices, and embedded profiles support interpretation of image data across systems. ACES separates a creative Look Transform from an Output Transform for a specific display and viewing condition; ACES Metadata Files can track input, look, and output transforms. Do not bake an unknown display look into source assets and call it portable. See the ICC profile-format introduction, ACES system overview, ACES Look Transforms, and ACES Output Transforms.
For each role—display, heading, body, caption, data label, supers, legal—specify:
Do not describe a typeface only as “bold sans.” Test the exact family at final pixel size, compression, motion, and viewing distance. Keep essential copy editable until final render; generation systems are not reliable typesetters.
For captions, reserve a predictable region, include speaker and non-speech information when required, and prevent important visual content from occupying that region. WCAG defines captions as synchronized equivalents for speech and needed non-speech audio and requires captions for prerecorded synchronized media at Level A; it requires audio description for prerecorded video at Level AA when visual information is not already conveyed. See WCAG 2.2 time-based media.
Define a small set of compositional behaviors rather than one frozen layout:
Specify recomposition, not cropping, for each required ratio. State which relationship survives when horizontal space becomes vertical space. For example: “In 9:16, preserve the subject-to-title diagonal and move proof points below the subject; do not center-stack every element.”
Production heuristic: Use one deliberate rule-breaking event per sequence. A centered, flat, or full-bleed exception reads as emphasis only when the baseline grammar is stable.
Name materials through observable properties:
Avoid contradictory bundles such as “matte mirror chrome” unless the contradiction is intentional and explained. Separate:
Set an imperfection budget. If every surface is distressed, nothing reads as intentional.
Describe:
Use diagrams or clock positions when useful. “Moody” is not a setup.
This is an example, not a universal recipe.
> One large cool key 35–50° camera-left and slightly above eye line; negative fill camera-right; one dim warm practical at least one subject-width behind; protect pale product labels from clipping; no frontal ring-light catchlight; keep shadow direction consistent across adjacent shots.
Describe the viewer–subject relationship first, then the optical means:
Documented fact: Shorter focal lengths give a wider angle of view and longer focal lengths a narrower one; keeping a subject the same size while changing focal length requires changing camera position, which changes the visible background and perspective impression. Sensor format also changes angle of view for the same focal length. Therefore record angle-of-view intent, subject distance, and sensor assumption—not a millimeter number alone. See Sony's first-party Lens Basics and Nikon's Understanding Focal Length.
This is an example for a calm technical product film.
For each motion class, specify purpose, object, path, duration range, easing character, overlap, sound relationship, and reduced-motion alternative.
Useful classes include:
Build a tempo map for the sequence. Identify rests, accents, accelerations, and the maximum number of simultaneous moving groups. Preserve object identity when transitioning between states.
Research finding: In two controlled experiments on transitions between statistical graphics, appropriately designed animated transitions improved graphical perception; the result is specific to related data graphics, not proof that animation always helps. Apply it when motion clarifies correspondence or change, not as a reason to animate decoration. See Heer and Robertson, “Animated Transitions in Statistical Data Graphics,” IEEE TVCG 13(6), 2007 (paper page).
Practitioner guidance with limited scope: Apple's interface guidelines advise purposeful, optional, brief motion and alternatives for motion-sensitive users. Material Design's published duration values are platform conventions, not universal film timing; use its deeper principle—scale duration with distance and complexity—without copying its milliseconds into unrelated media. See Apple Motion and Material duration and easing.
See WCAG 2.2 §2.3 and Apple's Reduced Motion evaluation criteria.
Create a continuity bible for every recurring entity:
Maintain a shot ledger with scene, shot, story time, entity state, screen direction, eye line, camera height, angle of view, focus target, lighting state, props, approved take/asset, and exceptions.
Documented fact: ScreenSkills describes script supervision as maintaining written and photographic records of action, costume, props, eye lines, camera, and lens details so shots made out of sequence can edit together consistently. Adapt that recordkeeping discipline to generated and hybrid media. See ScreenSkills: Script Supervisor.
Generate or shoot a continuity anchor set before the batch: neutral front/three-quarter/profile views, key expression or product states, material close-ups, lighting reference, and one difficult interaction. Reject a direction that cannot survive these tests.
Include:
Documented fact: W3C's image tutorial says text alternatives should convey an image's purpose, and complex graphics need a complete text equivalent of their data or information. See WAI Images Tutorial.
Do not claim that a contrast score alone makes a visual accessible. Test comprehension, sequencing, captions, audio description, and motion alternatives in context.
For every supplied or produced asset, record:
Documented fact: IPTC Photo Metadata provides descriptive, rights, and administrative fields, including copyright notice, rights-usage terms, instructions, credits, and digital-source information. C2PA Content Credentials provide a technical structure for recording media provenance and history; they establish provenance claims, not truth by themselves. The current C2PA specification shown by its official site was 2.4 when verified 2026-07-09. See IPTC Photo Metadata 2025.1 and C2PA specifications.
Use this minimum structure:
# Visual Direction: <project>
## Status
Version, owner, date, approval state, target deliverables
## Direction thesis
One sentence
## Brief interpretation
Audience, message, response, constraints, assumptions
## References and originality
Ledger; mechanisms adopted; identity details excluded; rights status
## Visual grammar
Invariants; controlled variables; exceptions; anti-patterns
## Systems
Color/contrast; typography; composition; texture/material; lighting;
camera/lens; motion; continuity; accessibility; color management
## Tokens
DTCG-compatible token block plus human-readable mapping
## Adaptation matrix
Still/video; 16:9, 9:16, 1:1, other targets; SDR/HDR; localization
## Proof set
Representative frames/shots, difficult cases, expected pass criteria
## Asset and rights ledger
Source, license, releases, credits, provenance, technical metadata
## Acceptance tests
Binary checks and subjective review questions
## Decision and repair log
Change, reason, affected rules/assets, approver
Add machine-readable JSON or YAML only if the production pipeline consumes it. Keep the prose packet as the semantic source of truth and ensure identifiers match across both.
Review at five gates.
Do not respond to “it feels off” with more adjectives. Locate the failing layer.
| Symptom | Likely cause | Repair order |
|---|---|---|
| Assets look individually good but unrelated | invariants too vague; too many uncontrolled variables | freeze 3–5 invariants; rebuild anchor set; regenerate/regrade outliers |
| Looks like the reference | surface/identity copied instead of mechanism | remove signature combination; introduce brief-derived composition, material, and type changes; rerun originality pass |
| Brand feels pasted on | logo treated as the only brand signal | move brand values into hierarchy, palette roles, material, voice, and motion; reduce logo dependence |
| “Cinematic” but empty | optical effects substituted for story hierarchy | restate communication beat; simplify camera; assign each move a reveal or relationship |
| Text passes on swatches but fails in video | dynamic background and compression ignored | test worst-case frames; add backing treatment or reposition; retest after encode |
| Batch continuity drifts | no canonical states or change ledger | build anchor views; assign entity IDs; lock shot ledger; repair highest-visibility contradictions first |
| Motion feels mechanical | identical timing/easing and no semantic classes | define reveal/orient/emphasize classes; vary by distance and importance; add rests |
| Direction collapses in 9:16 | cropping used instead of recomposition | preserve relationship hierarchy; define ratio-specific grid and negative-space ownership |
| Grade differs across systems | missing color-management contract | identify source/working/output spaces; restore transforms and profiles; reapprove on declared display |
Repair in this order:
Changing grain while the hierarchy is wrong is not a repair.
This is a complete example, not a mandatory formula.
Create a 30-second launch film plus six stills for a repairable modular desk lamp. Audience: design-conscious renters. Desired response: “This is engineered, calm, and made to last.” Deliverables: 16:9 hero, 9:16 cutdown, 1:1 stills. The client supplies a logo, CAD renders, product colors, one 1970s industrial catalog, one contemporary museum campaign, and macro photographs of anodized aluminum.
| Reference | Observed evidence | Transferable mechanism | Exclude |
|---|---|---|---|
| Industrial catalog | orthographic product views; dense part labels; warm off-white stock | evidence-led diagrams; numbered assembly logic; restrained warm ground | exact grid, typeface, illustrations, page damage |
| Museum campaign | one oversized object; wide negative space; title at edge | scale contrast and edge tension | signature crop, exact red/black palette, logo placement |
| Aluminum macros | directional brushing; cool highlight; soft falloff | material proof through grazing light | source photographs unless licensed |
“Turn repairability into quiet theater: precise diagram geometry meets tactile, softly worn materials, and every camera move reveals how one part connects to another.”
| Token role | Example sRGB | Use | Pairing test |
|---|---:|---|---|
| surface.mineral | #E9E3D7 | primary canvas | charcoal normal text ≥4.5:1 |
| surface.charcoal | #15191C | night/closing canvas | mineral normal text ≥4.5:1 |
| text.primary | #1D2327 | primary text on mineral | verify after texture/encode |
| accent.connection | #A84F35 | fasteners, active relationship, key label | never sole state cue; add shape/label |
| material.aluminum | measured/graded, not a flat swatch | product surface | preserve neutral highlights |
| Class | Purpose | Timing character | Reduced-motion version |
|---|---|---|---|
| Orient | move from whole lamp to joint | slow-in, decisive settle | cut from whole to joint with matched position |
| Reveal | expose replaceable fastener | short masked slide tied to part movement | two-state cut plus highlight |
| Compare | old module to new module | staged, one change at a time | side-by-side static comparison |
| Emphasize | confirm connection | brief copper line pulse; no flashing | persistent copper outline and label |
This is a complete example of a safe, useful redirection.
“Make my skincare campaign exactly like this living photographer's feed: same peach shadows, tight crops, wet glass, and serif.”
“Show clinical transparency through geological tactility: translucent product layers rest against carved mineral forms, with top light exposing ingredients and motion reserved for changes in layer or focus.”
The campaign preserves the requested intimacy and tactile evidence but has a different palette structure, material world, composition, typography, and camera behavior. It can be defended from the brief rather than from the photographer's identity.
Take calesthio/visual-style-direction from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.