--- name: creative-direction description: "Translate a brand's strategy — its emotional target and muses — into the concrete visual system: color, typography, composition, pattern & texture, iconography, illustration, and photography, as one coherent language. Use after brand-strategy and before logo or application work, or when the user mentions 'creative direction,' 'visual identity,' 'visual system,' 'design language,' 'stylescape,' 'pick our colors and fonts,' 'define our look,' 'present the directions,' 'compare the concepts,' 'mockups for each direction,' or an audit's moving-forward items. This is where 'what it should feel like' becomes 'what it looks like.'" metadata: version: 1.20.0 --- # Creative Direction You establish a brand's **visual system** — every asset class, one coherent language, all of it traceable to the feeling the brand is built to produce. This is the stage where strategy becomes visible: the emotional target and muses from `brand-strategy` get translated into actual color, type, composition, pattern, icons, illustration, and photography. Read `design-principles` first. This skill is where its rules get *applied*: contrast-not-competition, restraint, five stops, emotion before execution, and "state why, or the default stands." > **Position in the sequence:** after `brand` (the audit/context) and `brand-strategy` (the feeling + muses), before `logo-design` and the application skills. The audit *diagnoses*; this skill *builds the cure*. ## Hard Prerequisite: the Emotional Target and Muses Open `.agents/brand.md`. You need: - The **Emotional Target** (the feelings the brand must produce) - The **Muses**, each with its stealable visual cue - The **confirmed competitor set** (what "distinct" is measured against) - The **REO depth and the client's appetite** **If the Emotional Target or muses are missing, stop and route to `brand-strategy`.** Without them you'd be choosing colors and typefaces from taste — exactly the "different by accident" and "category default" failure modes the audit exists to catch. Every choice you make here must trace to a muse or the emotional target; if you can't say *why* a choice serves the feeling, it's a default wearing a decision's clothes. ## Scope by Depth - **Revolution** — build the system from scratch as **three genuinely distinct concepts**, presented as stylescapes (see [references/stylescapes.md](references/stylescapes.md)). Each concept is a different plausible answer to the emotional target — different muses, color story, type pairing, and attitude — not three shades of one idea. Stylescapes are *real intended assets*, not mood boards: expressive and free, not yet constrained by buildability. - **Evolution** — one direction, anchored on the **fixed core** (the equity being kept — often the logo, name, or a color). Start from the primary touchpoint (usually the hero/website) and change only the deltas the strategy named: carry the kept assets forward, introduce the new type, motif, or palette around them. - **Optimization** — mostly out of scope. The direction already exists; the work is execution (`web-design`, `collateral-design`). At most a light pass to *document* the existing system so it can be enforced. ## The Core Move: Abstract the Muses This is the engine of a distinct direction, and the step most likely to be skipped. **Take the muse's *literal* forms — recorded at physical fidelity in `brand-strategy` — and abstract them into modern, ownable brand elements.** You are not decorating with the muse; you are distilling its geometry, materials, and ideas into a system. - Worked examples (real): the concentric rings around the Kaaba → a dotted-line motif meaning *unity of people*; the building's brick geometry → an ownable grid; Islamic gardens → an "ants in the garden" pattern. The mission-control muse → the amber-on-black telemetry readout and the machined dimensional knob as the signature element. - **Pull two things from each muse: a color story and an abstracted idea.** Then combine the abstractions with modern UI. Most software looks the same — the muses are how you *skin* an undifferentiated category with something only this brand could own. - **Then render the *product's own work* through that abstraction — don't stop at decoration.** The single biggest gap is a board that ornaments the muse but never shows what the product *does*. Show the product's actual jobs — a post being drafted, a blog written, SEO run, a chart climbing up-and-to-the-right — expressed *in* the muse's language: musical notes that bloom into a published social post; a console readout that *is* the campaign launching. Muse ornaments floating over generic UI read as boring and unowned; the **muse × product-function merge** is what makes a direction feel like *this* brand doing *this* job. - **Don't let ease of abstraction crown the winner, and treat on-the-nose as a failure mode.** The muse that maps most literally to UI (a mission-control console → dials, telemetry) resolves fastest and looks the most finished soonest — but the literal reading is often the *weakest, most clichéd* idea, and cliché is a real liability in an AI category where everyone grabs the obvious. The muse that's *harder* to abstract (orchestration, empire) is frequently the more powerful, more ownable concept. Push harder on the difficult abstraction; don't retreat to the easy muse or fake the hard one with slop. Judge concepts on conceptual strength, not on which rendered first. - **Typography obeys the same abstraction law as everything else — and it has *two* failure poles.** A *literal* muse-face (an all-caps Roman inscriptional / Cinzel for an imperial muse) isn't a taste error, it's an **abstraction failure: you wore the muse instead of distilling it.** The opposite trap is just as fatal: reaching for a **try-hard trendy face** (Space Grotesk, Fraunces, and their cohort — **banned**; see `design-principles` → ai-slop.md) to *look* designed. Both are costume. The distinction must come from the **system** — motif, color, composition, treatment — not from the font. When in doubt, a **neutral face used deliberately** (Inter) beats a characterful one grabbed as a shortcut; let the system carry the muse. Muse-derived type is genuinely rare. ## The System to Establish Work through each element with the audit's bar in mind — the nine-category audit (see `brand`) is your acceptance test, and **every "Moving forward" item the audit raised (all nine categories except the logo, which `logo-design` handles after) must be answered by the system** — don't leave an audit gap unaddressed. Full per-element standards: [references/element-standards.md](references/element-standards.md). The production-craft floor — art-direction commitment, design tokens, typography/texture hard rules, the abomination checklist — lives in [references/frontend-boilerplate.md](references/frontend-boilerplate.md); the board's collage/bento composition in [references/stylescapes.md](references/stylescapes.md). 1. **Color** — assign by *role*: background, text (primary/secondary), **action/CTA** (protected — minimized everywhere else), support. Five stops per color, not ten. **Verify ownership against the confirmed competitor set** — a color isn't owned if a rival occupies it; deliberately escape the category's default lane. 2. **Typography** — faces by role (display, heading, body), chosen for **contrast, not competition**; if a mono or third register is needed, prefer a superfamily's own cut over a bolted-on face. Set weight/style restrictions. **Licensing checkpoint:** foundry, budget, and character/language support for the target markets — confirm *before* the direction is approved, not after. 3. **Composition & layout** — the grid, density, spacing logic, section rhythm and variety (no monotonous stacks), how the eye is directed. 4. **Pattern & texture** — an **ownable motif**, most naturally derived from the mark's geometry, deployable off-site without the logo or color present. Check it doesn't collide with a competitor's motif. 5. **Iconography** — one source of truth; stroke/fill/grid rules; plan the tiers (utility now, storytelling/illustration as the brand matures). 6. **Illustration & product abstraction** — branded abstractions over raw screenshots; a defined style (or an explicit "no illustration" decision). 7. **Photography** — subjects, treatment/grade, feel; or the deliberate absence of it. 8. **Motion** *(directional note, not full spec)* — what the brand's motion *feels* like; intent over ornament (hover > load > scroll > loop). Items 4–7 are decided here as *direction*; `art-direction` turns them into full specs and the actual assets. ## Producing the Board — Source and Generate the Real Assets A stylescape needs treated imagery and genuine texture; flat CSS shapes read as austere. The operator is often a **founder, not a designer**, so the craft leverage is on **you**. - **Get the imagery from `art-direction` — its board pass, once per concept.** It implies each concept's photography, illustration, icon, and texture style from the muses, finds web references the founder can point at, and produces the hero assets by the Toolkit's tier: generated directly with a connected tool, prompt packs for a generator the founder runs (ChatGPT, Gemini…), or free stock treated to the grade. It hands back treated assets; you build the board. Don't re-ask about tools — the Toolkit in `.agents/brand.md` already says what's available. - **You compose the board, not the founder.** They have no Figma skill and no design eye — so *you* assemble the coalesced board (self-contained HTML/CSS/SVG) from the generated + sourced assets. The operator supplies raw images (or you generate them); you do the art direction, layout, texture, and typography. - **Lean on CSS/SVG for the texture it renders well** — grain, noise, scanlines, patterns, grids, machined UI, telemetry, glows, gradients — which for many concepts (anything mechanical, retro, systems-y) is most of the "life," no generation needed. Reserve generation / free stock for photographic or illustrative muses. (Strict-CSP HTML can't load a webfont URL — embed the face as a data URI, or accept a fallback and *name* the intended typeface so the type intent reads.) - **Unify every source through one grade.** Whatever the origin — MCP-generated, founder-run, or free stock — run it through the concept's **treatment recipe** (a duotone + grain + scanline for a console; a warm-paper wash for a programme; a porphyry duotone for an imperial muse). Show it as a before→after on the board. The grade is what makes even free stock read as *this brand's* — and what keeps mixed sources looking like one world. - **The muses are the brief.** Everything generated or sourced is treated *to the muse cues at physical fidelity* — never generic stock dropped in raw. - **If imagery can't be produced yet, say so** and label the board a *composition / art-direction proof* with the prompt packs attached (for whichever generator the Toolkit names), so the founder can fill it in. Never quietly ship a flat board and call it finished. ## Validate Before You Present Run your own work through the same gauntlet the audit runs — **adversarially, on the rendered board, not the concept in your head** (see `design-principles` → Judge Your Own Work on the Competitor's Bar). The concept can be distinctive while the pixels are generic; a self-audit that grades intent misses exactly that. - **Score it against the foundational categories** vs. the confirmed competitor set, at the harshness you'd use on a rival — would this system earn 4–5 on Creative Direction, Color, Pattern, and (with `logo-design`) Logo? Cover the logo: is it still recognizably this brand? - **Run the AI-slop check** (`design-principles` → ai-slop.md) on your own output and **list what it finds** — the tropes actually on the board (a mono "telemetry/terminal" readout, a dot-grid, a left rail, a centered no-variety column, `rounded-lg` everywhere). "No slop here" is a reflex, not a result; produce the list, or an earned, specific "none, because —." "Clean" counts as *unassessed*, not passed. - **Confirm the board is a rendered surface, not a CSS-block comp** — treated imagery or a committed illustration style, a full named type specimen, an ownable motif (not a dot-grid default), applied mockups, real depth and bleed. Missing anatomy = under-built (see [references/stylescapes.md](references/stylescapes.md)). - **Carry-through gate:** the *applied* fragment (the hero/mockup) must pass cover-the-logo too, not just the abstract board (see `design-principles` → Distinctiveness Must Survive Application). A distinctive board with a generic applied screen is a fail — the distinctiveness has to survive into application, which is where it usually leaks out. - **Trace every element to a muse or the emotional target.** Anything you can't trace, cut or justify. - **Copy check:** does the system serve the message and give the words their proper emphasis? Design serves the message — the system is the stage, not the show. ## Present, Decide, Record - **Give it a walkthrough first, then present.** Talk each concept through — the muses, the abstraction, why it answers the feeling — before handing it over. A stylescape without its narrative is just arrangement. - **Name the challenges, honestly.** For each direction, say what would be *hard* to implement across the board, and **how different it actually is from the competitors** — the things the client needs to weigh to decide. Don't only sell; surface the tradeoffs. - **Close with a comparison — per direction or a synopsis at the end.** For each concept, answer the questions that actually drive the pick: the **challenges**, the **pros**, **is this the most differentiated?**, and **is this the most on-brand with the emotion you're trying to elicit?** Fold it into each direction's walkthrough or give a synopsis that weighs all three against each other at the close. - **Present in a presentation app (preferred).** One deployment per client: each direction shown as **Stylescape**, **Mockups**, and an interactive **Website**, with slider and side-by-side compare, comment pins, a Rationale panel (walkthrough, muses, thinking, board checklist), and a team verdict. A **shared copy deck** means every direction is judged on the same words. Structure and build: [references/presenting-directions.md](references/presenting-directions.md). - **Prove it has legs applied — on one fixed mockup kit.** Clients want to see it real, not abstract. Lock a standard set of scenes before any direction is built (laptop hero, phone, social/OG, email, business card, billboard, tee, plus a module for SaaS, physical product, or services), and apply every direction to the *same* scenes, angles, and copy, so the founder compares directions, not scenes. Kit, rules, sourcing, and checklist: [references/mockup-kit.md](references/mockup-kit.md). A concept that looks great on the board but thin when applied hasn't earned the pick (see the carry-through gate above). - **Document the styles and their sources.** Photo / graphic / icon / typography styles, and where any sourced assets came from (stock, Envato, Freepik, etc.) — notes now save the application skills from guessing later. - **Revolution:** present the three stylescapes; the client picks one. **Push for one idea, not a Frankenstein.** A little modularity is fine, but merging concepts usually averages three ideas into none. When a client asks to graft (e.g. this concept's type on that one's board), *knowing the emotional target lets you reject the merge that betrays it* — a merge is legitimate only if it still produces the feeling and has its own single story (see [references/stylescapes.md](references/stylescapes.md) → synthesis vs. averaging). - **Evolution:** present the direction against the *current* state so the deltas are unmissable. - Get **explicit approval** before `logo-design` and the application skills build on it — everything downstream compounds on this decision. (Run creative direction *before* any naming project.) - **Write the approved system into `.agents/brand.md`** (the Visual system section, plus updates to Fixed core / Flexible range), recording the *why* behind each choice — the rationale is what survives handoffs and lets a later critique judge flexes. ## The Deliverable Two layers, one source of truth — mirroring `brand-strategy`: 1. **`.agents/design.md`** — the full creative-direction record, generated at the end of this exercise. It holds: the three concepts and the one chosen; each concept's **muses → abstractions**; the documented **photo / graphic / icon / typography styles with their asset sources** (stock, Envato, Freepik…); the **color system** (roles + hex + ramps) and **design tokens**; the ownable **motif/texture** and the **signature element**; the **application examples** (hero + ads); the **implementation challenges** and **competitor-differentiation** notes; and the *why* behind every choice. Use the template in [references/design-doc-template.md](references/design-doc-template.md). This is the detailed spec the application skills (`web-design`, `collateral-design`, …) build against — the counterpart to `brand-strategy`'s `.agents/brand-strategy.md`. **And it's a living record:** downstream skills append their decided assets to it as the pipeline advances — `logo-design` the approved mark, `art-direction` the imagery specs, `motion-design` the built motion system, application skills their adopted libraries, notes, and a record per adopted component — what it is, how it's built, behavior at the extremes, use / don't use, what it must meet — and any skill its **Taste Rules**, owner-approved rules from the brand owner's corrections, per `design-principles` → Corrections Become Rules (see the template's Living Record sections). This skill writes the founding sections; the record grows from there. 2. **`.agents/brand.md`** — the distilled system: write the approved Visual system into its section (plus Fixed core / Flexible range) so every downstream skill inherits it without reading the whole record. Plus, if the client wants it, a **published copy** (Notion, PDF) generated *from* `.agents/design.md` — never the other way around. **Publish the system to Claude Design** once it's approved, if the operator uses it (Toolkit → design canvas): it becomes a design system that every new Claude Design project inherits, so anyone on the team can make on-brand decks, one-pagers, and prototypes without a terminal. Sync from the site repo if one exists, otherwise from a small kit in `.agents/assets/design-system/`; re-sync as `logo-design` and `art-direction` add to the record. Method, kit contents, and the sync checks: [references/claude-design.md](references/claude-design.md). Same rule as any published copy — `.agents/` stays the source of truth. ## Non-Negotiables - [ ] Emotional Target + muses read from `.agents/brand.md` (or routed to `brand-strategy` first) - [ ] Depth + appetite respected (three stylescapes / anchored deltas / documentation pass) - [ ] Color ownership verified against the confirmed competitor set - [ ] Type licensing + character support confirmed - [ ] Motif is ownable and doesn't collide with a competitor's - [ ] Every element traces to a muse or the emotional target - [ ] Adversarial self-audit run on the *rendered* board (AI-slop tropes listed, not "none"; applied fragment passes cover-the-logo) before presenting - [ ] Challenges + competitor-differentiation named; every direction applied to the same locked mockup kit with the shared copy deck - [ ] Client approval secured; full record written to `.agents/design.md`; distilled system written to `.agents/brand.md` - [ ] If the operator uses Claude Design: system published as a design system (kit or repo), validation checked, project recorded in `.agents/design.md` ## Related Skills - `design-principles` — the rules this skill applies - `brand` — the audit that diagnoses; its categories are this skill's acceptance test - `brand-strategy` — supplies the emotional target and muses (hard prerequisite) - `art-direction` — the imagery: photography, illustration, iconography, and pattern & texture, as specs and real assets (board pass for the stylescapes, system pass on the chosen direction) - `logo-design` — the mark, designed inside this system - `web-design` / `collateral-design` — apply the approved system with real copy - `design-critique` *(planned — not yet shipped)* — will judge later work against what's recorded here