Moodboard Generator

A moodboard is a single generated image that explores a brand's direction before anything is locked — a 3×3 grid of panels (color, type, logo sketches, icon style, photography, materials, pattern, mood imagery, one application mockup) that lets someone react to a feeling before committing to a hex code or a typeface. It's the format you reach for when the ask is "help me get oriented," not "give me the system" (brand-identity-sheet) or "give me the mark" (logo-prompting).

This skill is a sibling to brand-identity-sheet and logo-prompting — same two-artifact workflow, same Ideogram generation pattern, same reference-file structure — but it sits earlier in the process than both of them. Where those two skills expect a brand.md with real decisions already made, a moodboard is often the thing that produces the first draft of those decisions. Tolerate fuzzy, partial, or three-word inputs here in a way the other two skills shouldn't.

The two artifacts this skill produces

1. A generation prompt — a tight descriptive paragraph, Ideogram-legible, that produces the 3×3 board in one shot.
2. A compositional deconstruction — a structured JSON breakdown of the resulting image (or of a reference image the user supplies) with a high_level_description and an elements array of {type, bbox, desc} / {type: "text", bbox, text, desc} entries. Same schema brand-identity-sheet and character-model-sheet use, unmodified — see references/composition-spec-format.md. This is what lets a later brand-identity-sheet or logo-prompting pass point back at "the logo-exploration panel from the moodboard" as a starting reference instead of re-describing it from scratch.

Always produce both. The paragraph prompt is what you feed the image model; the JSON is what you save back to the vault so the board's structure survives past the single image.

Workflow

1. Pull whatever brand truth exists — don't block on missing pieces

If the project has a settled position or a token lock, read them first. The position comes from position-recommender and tokens from kindform-tokens (DESIGN.md); treat both as settled. When they exist, the direction is already made and your job is to execute it, not to re-open it. Re-deciding by eye is how a derived token set quietly becomes taste again.

What to take, and from where:

A project with an older position.yaml and no DESIGN.md: read the same five things from it (tokens.colour, tokens.typography, diagnosis, intent, selection.fails_at).

Leave confidence labels and source appendices alone; they belong to the position document, not to this artifact. And if a settled file disagrees with something the user just told you, say so out loud rather than silently picking one. A contradiction between a standing position and a fresh instruction is a decision for the user, and it usually means a re-run (position-recommender for the position, kindform-tokens for tokens) rather than a workaround here.

Where neither exists, carry on below as normal.

Read the project's brand.md if one exists. But unlike brand-identity-sheet, a missing or half-written brand.md is the normal case here, not a blocker — a moodboard is frequently the first artifact in a project, run specifically to help someone figure out what the brand.md should say. Gather what you can, invent nothing, and leave the rest as an open exploration:

• Adjectives — three words for the brand's personality. If the user hasn't given these, ask for them directly; this is the one input worth pausing for, since without it every panel defaults to generic. If they truly don't have words yet, pull from whatever context exists (industry, name, one sentence about who it's for) and propose three candidates back to them rather than inventing silently.
• Color focus — a hex, a named hue, or even just "warm neutrals" / "cool and clinical." Open palette is fine — a moodboard's whole point is exploring this.
• Logo concept — a symbol or motif the user envisions, if any. Often blank at this stage, and that's fine — panel 3 (Logo Exploration) can sketch possibilities rather than lock one.
• Imagery & texture — preferred visual/material references (a name-brand aesthetic, a material like "brushed steel" or "paper grain," a mood photographer's work).
• Emotional impact — one feeling the whole board should evoke (calm, urgent, playful, trustworthy).
• Notes — anything else that doesn't fit the above (a competitor to avoid looking like, an industry constraint, a reference image).
• Human presence — if the brand's claim runs through people at all (trust, care, service, who it's for, who it's by), name who should appear in the Photography & Graphics and Mood Imagery panels, and from what vantage point: their eye level, not looking down at them, not a stock-photo wide shot. If the claim doesn't involve people, say so explicitly rather than leaving it unaddressed. An unaddressed people-question defaults to "nobody" or to the wrong somebody, and neither is a neutral choice. This belongs in Notes and Emotional Impact rather than as its own template variable, but it has to be asked, the way adjectives are asked, because the failure it prevents is a question nobody raised, not a field nobody filled.

These six map directly onto the template variables in references/panel-anatomy.md. If a project already has a brand.md, pull adjectives from its Voice layer and color/imagery from its Visual layer rather than re-asking — same as brand-identity-sheet does — but don't require it to exist first.

2. Fill the 9-panel template

The panel set is fixed for this skill (unlike brand-identity-sheet's pick-and-choose panel list) — the whole point of a moodboard is breadth across all nine categories in one board, not a curated subset. See references/panel-anatomy.md for what each panel is actually testing and what makes it earn its place rather than feel decorative:

1. Color Palette
2. Typography
3. Logo Exploration
4. Iconography Style
5. Photography & Graphics
6. Material Samples
7. Abstract Patterns
8. Mood Imagery
9. Application Mockup

3. Write the paragraph prompt

Structure, in order — same discipline as the sibling skills, applied to a 3×3 exploration board instead of a locked system:

1. Frame it as a brand moodboard / visual exploration board, arranged as a 3×3 grid — naming the grid explicitly and the word "exploration" is what keeps this reading as an intentional research board rather than a single confused collage, the opposite problem brand-identity-sheet solves by avoiding the word "moodboard" entirely (that skill wants the locked-system look; this skill wants the exploration look, so lean into it here).
2. Name all nine panels in reading order (left-to-right, top-to-bottom), each with a short label so Ideogram renders them as nine distinct cards, not nine blended scenes.
3. Adjectives: state the three words plainly — they're the thread that should be legible across every panel, not just the mood-imagery one.
4. Color: whatever's known — exact hex if locked, a named hue or temperature if still open ("warm terracotta and cream, still exploring the exact palette"). A hex alone is not a strong enough anchor for a light, warm, low-saturation ground. Ideogram drifted #E8E4DA to pink-beige and then to cream-yellow across two independently drafted prompts on the same brief. For a ground in that register, anchor it twice: state the hex, and in the same clause name what it is not ("a warm paper neutral, not pink, not yellow, not cream"). Saturated colours and dark grounds held accurately in the same run, so this is a narrow rule, not a blanket one.
5. Logo concept: describe it as a sketch or exploration, not a finished mark — "a loose sketched exploration suggesting X," never "the final logo is X." This is the one place this skill deliberately contradicts logo-prompting's "abstract, don't depict literally" rule less strictly — a moodboard's logo panel is allowed to be more literal and rougher, because its job is to spark a direction, not ship a mark.
6. Imagery & texture, material samples, abstract patterns: pull straight from whatever references the user gave — name real materials and real reference aesthetics rather than "some textures."
7. Emotional impact: one sentence naming the feeling the whole board should leave someone with.
8. Negative space / anti-slop: see references/anti-slop-discipline.md — a moodboard with nine panels has nine chances to drift into generic AI-image territory, more than either sibling skill.
9. Background/composition note: clean white background, thin grid lines separating the nine panels. Ask for small panel captions if a labeled board reads more like an intentional exploration than a collage, which is what keeps this from looking like a Pinterest dump. But caption text at that scale reliably comes back garbled; Ideogram cannot set small type legibly across nine panels in one render. Treat a caption as compositional furniture that makes the board look structured, never as text to be read. If the board genuinely needs readable labels, add them afterward in an image editor, or skip captions and let panel position (reading order, left to right, top to bottom) carry which-panel-is-which.
10. What the Logo Exploration and Typography panels can and cannot tell you. Ideogram cannot letter reliably. It renders shapes that resemble type, not a typeface. Four boards on one brief returned four different invented faces, none of them the real typeface already locked for that brand. Read those panels for composition and mood only: heavy or light, formal or loose, never which face rendered. Say this plainly in the saved file next to the images rather than leaving it implied. If the board runs before any typeface is locked, the panel is genuinely exploratory and the caveat does not apply yet; it exists for the commoner case where a real typeface exists and a generated board could be mistaken for challenging it.

Keep it one paragraph. If the user wants to compare genuinely different directions (not just a palette swap), write separate prompts — same "structural variety over palette-swaps" rule as logo-prompting and brand-identity-sheet: two boards that only swap which hex fills the same nine slots aren't two directions.

3.5. Run the anti-slop gate before generating

Before calling generate_image, score the drafted prompt against the pre-generation gate table in references/anti-slop-discipline.md. Anything scoring under 3 on any axis means revise the prompt now — a nine-panel image is expensive to regenerate, and a loose "vibe board" that pulls Ideogram toward generic Pinterest-collage energy is the single most common failure mode this skill exists to prevent.

4. Generate it — don't just hand back a prompt

Call mcp__ideogram__generate_image with the paragraph prompt. Use style_type: "DESIGN" and aspect_ratio: "1x1" — a 3×3 grid reads cleanest as a square. Use rendering_speed: "QUALITY" for the same reason the sibling skills do: nine text-bearing panels at once is a composition-heavy, legibility-sensitive render.

If the user supplies a reference moodboard they like and wants variations, use mcp__ideogram__remix_image instead of a fresh generate_image call — image_weight around 30-50 for "same energy, different brand," since moodboards are meant to be loosely interpreted rather than precisely reproduced (contrast with brand-identity-sheet's tighter 40-60 remix range for a locked system).

4.5. Run the post-generation check before treating the image as usable

A prompt that scores well on the pre-generation gate can still come back structurally wrong. The gate measures what you asked for, not what Ideogram built. One board on this brief scored 5/5/4/5/5 going in, then collapsed from nine panels to six, rendered the wordmark as a script face, and substituted an illustrated typewriter where a documentary photograph was specified. Nothing caught it, because the family stops at score-prompt, generate, save.

Before writing anything up as usable, look at the returned image against references/post-generation-checklist.md. A board that fails gets marked discard in the saved file, exactly as a board failing the pre-generation gate would have been revised before generating. The only difference is which side of the render the failure showed up on.

5. Reverse-engineer into a compositional deconstruction

Once the board is generated (or if the user hands you a reference moodboard to work from), produce the structured JSON breakdown described in references/composition-spec-format.md — same process as the sibling skills:

• From a generated image: write the JSON directly from the panel layout you specified in step 3.
• From a reference image: call mcp__ideogram__describe_image first, then structure that description into the schema, using coarse-but-honest bbox estimates where you can't confidently place exact coordinates.

6. Save what you made

Before saving, see references/ideogram-asset-retrieval.md for how to get the real image file rather than the permalink page, and for the WebP conversion macOS needs before the file can be viewed at all.

Write the paragraph prompt, the JSON compositional deconstruction, and the resulting image URL(s) to 04-projects/<project>/logo-explorations/moodboard-<label>.md (or wherever that project keeps its branding work — check for an existing logo-explorations/ or branding/ folder first and match it). Follow the "No Context Lost" rule: nothing generated in this skill should live only in the conversation.

If the user reacts strongly to a specific panel (loves the palette, hates the logo sketch), capture that reaction in the saved file explicitly — it's the input the next brand-identity-sheet or logo-prompting pass will need, and a moodboard's whole purpose is feeding that decision forward.

When to stop generating and write a brief instead

Some of what a moodboard is asked to settle, a photographic register, a typographic feel, an illustration style, cannot be delivered by Ideogram at final quality. It can only be gestured at. Watch for the specific signal that a board has crossed from useful exploration into the wrong medium for what is actually needed now: the render answers the casting and framing question convincingly (who is in the shot, what mood, what composition) while failing the register question (is it really a photograph, is it really set in a real typeface) for structural reasons another generation cannot fix.

When both are true at once, the next deliverable is a brief for a human photographer, illustrator or type designer, written from what the board settled. Not another round pushing the same image closer to a photograph it structurally cannot become. Say it explicitly in the saved file: name which panel is now a brief, and who it is for.

The judgment is real and cuts both ways. Declaring "brief" before genuinely trying to iterate the prompt wastes the cheap medium; declaring it after four wasted rounds wastes the expensive one. The signal above is the thing to key off, not a general sense that the image is disappointing.

Reference files

• references/panel-anatomy.md — the nine named panels (Color Palette, Typography, Logo Exploration, Iconography Style, Photography & Graphics, Material Samples, Abstract Patterns, Mood Imagery, Application Mockup), what each one is testing, and what makes it earn its place rather than feel decorative.
• references/composition-spec-format.md — the JSON schema for the compositional deconstruction (shared, unmodified, with brand-identity-sheet's and character-model-sheet's schema).
• references/anti-slop-discipline.md — the moodboard-specific anti-slop guardrails (nine panels of surface area, the "exploration vs. collage" line, the pre-generation gate) adapted from brand-identity-sheet's version.
• examples/ — worked moodboard jobs go here as they happen; empty for now.