Art Direction

Art Direction

Build a defensible visual identity from approved strategy, character, verbal identity, competitive context, semiotics, and human-curated references. The output is a visual system capable of producing real surfaces, not a moodboard or a pile of generated assets.

The human owns the frame, taste, territory choice, and final quality bar. Agents can audit, organize, generate, test, and repeat the system, but they cannot infer the intended world from generic best practice.

When to invoke

• When creating or rebuilding a visual identity, event identity, digital product, or master-brand system
• When a team needs visual territories, competitive differentiation, or reference guidance
• When a moodboard feels attractive but arbitrary, generic, or hard to extend
• When defining typography, color, composition, image, motion, material, or interface rules
• When deciding whether commissioned, stock, or AI-generated imagery belongs in a brand world
• When translating an approved identity into Figma components, templates, or repeatable assets
• When critiquing visual work against the intended concept rather than personal preference

Load the system

Read:

1. ../../foundation/marketing-os/SKILL.md
2. ../../foundation/marketing-os/references/method.md
3. ../../foundation/marketing-os/references/collaboration.md
4. ../../positioning/{relevant brand}/SKILL.md
5. ../../brand-voice/{relevant brand}/SKILL.md
6. ../../marketing-science/research/SKILL.md
7. ../../marketing-science/archetyping/SKILL.md

Load ../../strategy/compound-brand/SKILL.md and ../../craft/verbal-identity/SKILL.md when art direction is part of a full identity workflow. Load brand equity only for an explicit evaluation. For a formatted collaborative artifact, also read ../../foundation/marketing-os/references/native-document-editing.md.

Reference routing

Establish authority

Identify the sources that own:

• Strategy, positioning, audience, character, and verbal identity
• Current logos, typefaces, colors, variables, components, and templates
• Approved references and anti-references
• Product, channel, format, accessibility, and production constraints
• Final visual approval

Inspect the actual design files and assets before proposing a system or automating one. Distinguish the canonical system from examples, experiments, exports, and stale duplicates. Do not silently replace unavailable fonts, missing assets, or unresolved decisions.

Require the human reference gate

Require five to 10 human-supplied references and at least three anti-references. Use references/reference-doctrine.md to guide the selection. Do not proceed from generic labels such as Bauhaus, vintage editorial, graffiti, or surrealism.

For each reference, identify what formal rule it contributes, what it means in this context, why it is fresh in the category, and what must not be copied literally. The human can revise the set after analysis; the gate exists to expose taste, not to freeze an early moodboard.

Audit the field

Load references/competitive-semiotic-audit.md. Audit competitors' marks, typography, color, composition, imagery, motion, materials, interfaces, and category codes. Return a fit-in versus stand-out diagnosis. Recommend standing out unless category legibility, trust, safety, distribution, or an intentional dupe strategy requires borrowed codes.

Develop territories

Offer no more than two:

1. Functional-led: Organized around what the brand demonstrates or helps people do.
2. Emotional-led: Organized around the feeling, identity, or cultural tension it owns.

Each territory must combine at least two specific reference domains in a way that feels new in the category. Explain the conceptual bridge. Present the governing idea, reference logic, semiotic codes, system grammar, risks, anti-references, and representative surface behavior.

The human chooses, rejects, or edits the territory before asset development. Do not hide an unresolved strategic choice inside visual polish.

Define the visual identity system

Load references/visual-identity-system.md. Define:

• Brandmark and wordmark
• Type hierarchy
• Color system
• Layout and compositional grammar
• Image, illustration, and photography direction
• Glyphs, sigils, or icons when needed
• Texture, material, and brand-world elements
• Motion behavior
• Digital product and website expression
• Print, OOH, merchandise, environmental, and accessibility behavior

Describe how the system behaves before discussing minute applications. A badge, headshot treatment, social card, or token can prove the system later; it should not substitute for the system.

Make the system operable

Load references/agent-readable-production.md before building components, templates, or automated asset workflows.

Treat design-system semantics as production infrastructure. Use stable, human-readable names for components, variants, properties, frames, fields, and source assets. Prefer a full identifying name when two people or objects share a short label. Name by identity and role, not by coordinates, appearance, or temporary order.

Define which environment owns each responsibility. Figma may own layout, geometry, variables, and canonical components while another renderer owns typography or export when tooling cannot reproduce custom fonts reliably. Record the boundary instead of pretending one tool owns the entire system.

Separate four modes of work:

1. Diagnose: Inventory the real state and find structural mismatches.
2. Systemize: Establish canonical components, variables, names, and constraints.
3. Produce: Generate the requested surfaces from those rules.
4. Verify: Inspect structure and rendered output independently.

Automate repetition. Escalate ambiguous crops, image selection, optical spacing, and other taste judgments to a human when correction is cheap. Record the correction so the next run improves.

Direct asset generation

Specify whether an asset should be commissioned, photographed, illustrated, designed, sourced, or generated. Name the production method, tool, human role, iteration loop, rights constraints, and quality-control standard. Generated imagery must share recurring rules and survive human art direction and postproduction. Raw generation is never an approved final merely because it is technically complete.

For generative exploration, load references/agent-readable-production.md. Lock the approved content, exact source surface, reference roles, and acceptance criterion before branching. Begin with one real surface, preserve the strongest result, and change one named variable per branch. A generator can execute a visual decision; it should not be asked to invent the art direction and evaluate itself in the same pass.

Prove and evaluate the system

Test a small, representative set of surfaces before scaling. Include at least one dense, one sparse, one small-format, and one large-format application when relevant. Verify:

• Argument: Choices trace back to the strategy and governing idea.
• Distinctiveness: The work is recognizable without relying only on a logo.
• Category freshness: It does not reproduce category wallpaper.
• System value: The idea can generate varied surfaces without becoming arbitrary.
• Operational fitness: Names, components, assets, constraints, and ownership are clear.
• Accessibility: Type, contrast, motion, and interaction survive real use.
• Rendered quality: Crop, scale, overflow, spacing, and typography hold in final output.

Use both structural checks and visual proof. A correct component tree can still produce a bad crop; a good screenshot can conceal the wrong identity, field, dependency, or export path.

Output

Right-size the work, but preserve this order:

1. Source and constraint inventory
2. Competitive and semiotic audit
3. Reference and anti-reference analysis
4. One or two territories
5. Human territory decision
6. Visual identity system
7. Representative surface proofs
8. Production architecture and naming rules
9. QA findings, open decisions, and approval owner

For a critique, report the governing idea first, then what strengthens it, what breaks it, and the smallest decisive correction. Do not replace a point of view with a long list of equally weighted options.

Dependencies

• foundation/marketing-os
• strategy/compound-brand for a full identity workflow
• craft/verbal-identity for a full identity workflow
• marketing-science/research
• marketing-science/archetyping
• marketing-science/brand-equity when explicitly evaluating equity
• positioning/{relevant brand}
• brand-voice/{relevant brand}

Quick checklist

• Strategy, character, verbal identity, product, and channel context are explicit
• Canonical files, components, assets, fonts, variables, and approval owners were found
• The human supplied five to 10 references and at least three anti-references
• Competitive visual and semiotic audit is complete
• References are specific, category-fresh, and system-generating
• No more than two distinct territories were presented
• The human locked the territory before asset development
• Every relevant visual identity component is defined
• Components and assets use stable semantic labels
• Production ownership and tool boundaries are explicit
• Representative surfaces passed structural and visual QA
• Human corrections were captured as reusable rules where possible