Art Direction

You make the brand's imagery — every picture, drawing, icon, and surface it shows — look like it came from one world, and that world is the one the strategy described. creative-direction decides the direction; this skill turns it into four concrete styles and the assets themselves:

• Photography — what's in frame, how it's lit, shot, and graded
• Illustration — the drawn language, including branded product abstractions
• Iconography — the symbol set and its rules
• Pattern & texture — the motif's deployments and the surface library
• …and the combination — how the classes coexist on one surface

Read design-principles first. Everything here is its rules applied to pixels: emotion before execution, restraint, coherence not repetition, and "state why, or the default stands."

Position in the sequence: an extension of creative-direction. It runs inside it for the stylescapes (the board pass), again on the approved direction (the system pass — before or after logo-design), and any time an application skill needs imagery the library doesn't have yet.

Prerequisites

Open .agents/brand.md and, for a system pass, .agents/design.md. You need:

• The Emotional Target and the Muses with their stealable cues — if missing, stop and route to brand-strategy. An image style chosen without them is stock-photo taste.
• For a system pass, the chosen direction in .agents/design.md — if there isn't one, route to creative-direction. Don't art-direct a system that hasn't been picked.
• The Toolkit — what the operator can generate with, and their tier. If it's missing, run the capture step from brand (detect what's connected; ask once about tools they run by hand) and write it back.

Two Depths

• Board pass (inside creative-direction, once per concept) — enough to make each stylescape real: the four styles in a paragraph each, a reference set, a quick style test, and a handful of hero assets produced and treated. Breadth across concepts, not a library.
• System pass (the approved direction) — the full spec per class, the anchor set, the treatment recipe, the combination rules, and an approved library the application skills draw from.

Step 1 — Imply the Styles From the Strategy

Before you search or generate anything, write each style in plain words, derived from the emotional target and the muses — the same abstract the muse move creative-direction makes, applied to imagery. For each class, one short paragraph a non-designer could picture:

Photography: "Working hands and materials, close and unposed. Hard window light from one side, deep shadow, nothing lit evenly. Warm highlights, cool blacks, visible grain. Never a face looking at the camera."

• Trace every sentence to a muse cue or a feeling. "Hard side light" because the muse is a machinist's workshop at dusk; "never a face to camera" because the target is quiet competence, not friendly. Untraceable lines are defaults — cut them.
• "None" is a valid answer, stated explicitly. A brand with no photography or no illustration says so, with the reason — otherwise every future deck fills the vacuum with mismatched stock (see creative-direction → element-standards.md).
• Name what the category does, and step out of it. If every competitor runs bright, diverse-team-around-a-laptop stock, that's the lane you're leaving — say so in the paragraph.

Step 2 — Show, Don't Describe: Find References on the Web

A paragraph about light is abstract until the user can point at it. Go find real images that fit each style and put them in front of the user — "Based on your strategy, here are a few photos I found that have the feeling we're after." Full method, sources per class, and the board format: references/sourcing-references.md.

• 4–6 per class, each captioned with the one quality it contributes — "the light," "the crop," "the grain," "how the figure sits in the space." A reference is a pointer to a specific quality, not a vote for the whole image.
• Include one deliberate near-miss per class and say why it's close but wrong. Where the user draws that edge tells you more than ten approvals.
• Let them point. Record which images landed, which didn't, and what about them — then revise the paragraph from Step 1 to match what they actually responded to.
• Mark every reference reference only or usable (license). Unsplash and Pexels images and public-domain museum collections can ship; a photographer's portfolio or a Behance piece cannot. Reference-only images never appear in a deliverable.
• Icons and texture are rendered, not just found. For icons, render the brand's own concepts in 2–3 candidate libraries side by side at real size. For texture, render CSS/SVG candidates directly next to the material references.

Step 3 — Write the Spec

Turn the agreed paragraph and the picked references into a spec tight enough that two people working separately would produce images that look like siblings. Field-by-field templates per class, plus the combination rules: references/style-specs.md.

• Specifics, not adjectives. "Moody" is not a spec; "single hard source at 45°, shadows held at 10–15% luminance, no fill light" is.
• Palette anchors in hex, from the system's color roles. The action color stays out of imagery unless the spec says exactly where it may appear.
• A never list per class — the things that would instantly pull an image out of the brand world. It's often the most useful part of the spec.
• Combination rules — which classes share a surface, what layers over what, how icons relate to the illustration's line, where texture unifies everything, and which class leads in which medium.
• The treatment recipe — the one grade every image runs through, whatever its source. Write it as values (duotone stops, grain amount, contrast curve), and implement it once in code: an ImageMagick or sharp script for files, an SVG filter or CSS for the web. Show it as a before→after.

Step 4 — Test the Style Before You Lock It

A spec that reads well can still fall apart the moment it meets a second subject. Prove it on a test sheet before anything is built on it — illustration above all, because it's the class where the same words produce the widest spread. Sheet format and the subjects to include: references/style-specs.md → The Style Test.

• Same spec, deliberately different subjects. For illustration: a person, an object, a place, an abstract idea the brand actually has to explain, and the product as a branded abstraction — at hero size and at spot size. Photography varies subject and setting while holding light and grade; icons run the brand's concept set at every size tier; texture runs at three scales behind real type.
• Include the hard cases — hands and faces, crowded scenes, abstract concepts, the smallest size. That's where a style breaks.
• Run it twice. The same prompts again (or in a second tool): if the two sheets aren't siblings, the spec is underspecified. Find the field that drifted — line weight, color count, perspective, detail — and tighten it.
• Do what you can yourself. Line-led, flat, and geometric styles render well in SVG, so you can build a first sheet in code; with a connected tool (Flora, Paper, Figma, Adobe), generate the sheet in-session — Arrow (via Quiver, Flora, or Paper) produces flat illustration and spot icons as real SVG. Painterly, textured, or dimensional styles need a generator: at Tier 1 the test sheet is the first prompt pack. Much of the heavier testing happens in those tools — run it there when they're available.
• Approve the style on the sheet, not the spec. Show the user the sheet, in place (a section, a post, a slide) and next to the references they picked. Update the spec with whatever the test taught you; the sheet's winners become anchor candidates. Save it to .agents/assets/imagery/tests/.

Step 5 — Pick the Anchors

Choose 3–6 anchor images per class — the best embodiments of the spec, from usable references or approved generations. They are the brand's visual memory: every later generation is made with the anchors attached as references, which is what keeps image #40 looking like image #1. Store them in .agents/assets/imagery/anchors/.

Step 6 — Produce, by Tier

Read the Toolkit and take the most capable path available. Full playbooks — the ChatGPT Project / Gemini Gem primer, prompt anatomy per class, edit-first iteration, and connected-tool notes: references/generation.md.

• Tier 2 — a generator connected to you (Flora, Adobe Firefly, Higgsfield…): generate directly, anchors attached as style references. Where the tool can train a style model on the anchors, do it once the anchor set is approved — it's the strongest consistency lever there is.
• Tier 1 — a generator the operator runs (ChatGPT, Gemini, Midjourney): write a primer — one block of instructions for a ChatGPT Project or Gemini Gem that carries the whole spec — so each request becomes one line plus 2–3 attached anchors. Save the primer and every prompt pack to .agents/assets/imagery/prompts.md for reuse. The operator saves results to .agents/assets/imagery/inbox/; you take it from there.
• Tier 0 — Claude only: usable free stock (Unsplash, Pexels, public-domain collections) chosen against the spec and run through the treatment; icons from the chosen library; illustration and texture in SVG/CSS where the style allows. Say plainly that generated photography isn't available at this tier rather than faking it.

Edit before you regenerate. When an image is 80% right, ask for the specific fix ("keep everything; drop the second figure; cool the shadows") — current models edit in place. Regenerating throws away what was working.

Step 7 — Review Every Image Against the Spec

Nothing enters the library unreviewed. Look at each image — actually look — and check it rule by rule against its class spec and the never list, then against references/imagery-slop.md — the tells that mark an image as generic or AI-made.

• List what you find, per image. "Looks on-brand" is a reflex, not a review; "clean" counts as unassessed.
• Give one verdict per image: approve · edit (with the exact edit prompt) · re-prompt (what to change in the prompt and why) · reject.
• Judge it in place, not alone. Drop the candidate into the surface it's for — the hero, the post, the slide — next to the approved set. An image can pass on its own and still break the family.

Step 8 — Treat, File, Record

• Run every approved image through the treatment recipe. Mixed sources — generated, stock, shot — only read as one world after one grade.
• File it: .agents/assets/imagery/approved/<photography|illustration|icons|texture>/, and add a line to .agents/assets/imagery/manifest.md — file, class, source or tool, prompt (or source URL), rights, alt text, date, intended use.
• Record the spec as the Art Direction section of .agents/design.md (living record — see its template), and the distilled lines into the Photography, Illustration, Iconography, and Pattern & Texture fields of .agents/brand.md.

Rules That Don't Bend

• No logo and no brand type inside generated imagery. Generators garble both, and a baked-in mark can't be reused or updated. Set the type and place the locked logo files (.agents/assets/logo/) on top, in layout.
• Never prompt "in the style of [a living artist or photographer]." Describe the qualities you want — light, line, palette, texture. It's both the ethical line and the better prompt.
• Never fabricate the real. No generated "customers," "team members," or testimonial portraits presented as real people; no generated likeness of a real person; no generated product shots passed off as the actual product; no generated "archival" or historical photograph — when the muse is history, use the real archive (much of it is public domain). Show the product through real screenshots or branded abstractions (see creative-direction → element-standards.md).
• Reference-only never ships. If the right image is a reference you can't license, generate or shoot toward it with the anchors — don't use it.
• Rights are recorded, not assumed. Note each tool's commercial-use terms and each stock license in the manifest.

The Deliverable

1. The Art Direction section in .agents/design.md — the four specs, combination rules, treatment recipe, anchor list, and the tool notes (which generator, which primer).
2. The distilled lines in .agents/brand.md — Photography, Illustration, Iconography, Pattern & Texture.
3. The library — .agents/assets/imagery/ with anchors/, approved/, manifest.md, the reference board and its source list in references/, the style test sheets in tests/, and prompts.md at Tier 1.
4. For a board pass: the produced, treated hero assets handed back to creative-direction for the stylescapes.

Non-Negotiables

• Emotional target, muses, and (system pass) the chosen direction read first — or routed upstream
• Toolkit read; production path matches the tier, and the tier is stated honestly
• Each style written in plain words and traced to a muse or the feeling; "none" stated explicitly where it applies
• Web references shown per class, captioned, with a near-miss; picks recorded; every reference marked reference-only or usable
• Spec per class with hex anchors, a never list, combination rules, and one treatment recipe
• Style tested on a sheet — varied subjects, hard cases, sizes, run twice — and approved on the sheet (illustration always)
• Every image reviewed against its spec and the imagery-slop list — findings listed, one verdict each, judged in place
• No logo or brand type in generated images; no fabricated people or product; no reference-only image shipped
• Library filed with manifest (rights + alt text); spec written to .agents/design.md, distilled to .agents/brand.md

Related Skills

• design-principles — the craft beliefs every spec traces to
• brand-strategy — supplies the emotional target and the muses (hard prerequisite)
• creative-direction — decides the direction and the motif; calls this skill for the board pass; its element-standards.md sets the bar each class must clear
• logo-design — the locked mark placed over imagery, never generated into it; its geometry seeds the pattern
• web-design / collateral-design — draw imagery from the approved library; route back here for anything new
• motion-design — motion character for any generated video or animated asset

Worked example: examples/magister — a full run, including the style test with every verdict (.agents/assets/imagery/tests/style-test.html).