Concept Grid Generator

Concept Grid Generator

A concept grid is a single generated image that puts several genuinely distinct motif/icon concepts side by side as labeled tiles, with the palette and the rendering technique held identical across every tile. It exists to make choosing between candidate ideas cheap: one generation, one review pass, instead of running a full brand-identity-sheet or character-model-sheet for every candidate and comparing finished boards after the fact.

Why this exists — cheap N-distinct-concepts comparison vs. N expensive full boards

A full identity sheet or character sheet is a composition-heavy, multi-panel, QUALITY-rendered generation — it's the right tool once a direction is picked, but it's the wrong tool for picking between directions. Running three full boards to compare three candidate motifs triples the cost and the review burden for a decision that doesn't need six-panel context, it needs to see the marks next to each other. A concept grid answers "which of these ideas is worth developing" in one generation by holding everything except the concept itself constant, so the only variable a reviewer has to judge is the idea.

This is different from varying a locked idea (bulk-image-generation-workflow, which holds style_description byte-for-byte identical and varies pose/prop/angle within one concept) and different from exploring a whole brand loosely (moodboard-generator, which explores nine categories — palette, type, imagery, mood — with the palette itself still open). Here, the palette and the rendering technique are the two things locked before drafting a single tile, and what varies is the concept — the actual form/idea each tile represents.

The two-stage handoff — Set A (this skill) → Set B (lock-in)

Treat this skill as Set A and brand-identity-sheet / character-model-sheet as Set B. Never skip straight to a full board for every candidate concept — that's the expensive path this skill exists to avoid. And never treat the grid image itself as a finished deliverable — a tile in a comparison sheet is a rough, small, same-technique sketch next to five others; it hasn't been rendered at the scale, polish, or system-context a shipped identity sheet or character sheet needs.

The handoff, in order:

1. Run this skill first. Produce the grid, review it with whoever's deciding, and get

an explicit pick — one winning concept, named.

1. Route only the winner forward, as a separate follow-up generation, into

brand-identity-sheet (if the winner is a brand mark/icon direction) or character-model-sheet (if the winner is a character/mascot direction). That skill develops the winning concept into its own full panel set, typically with the same locked palette carried forward and the rendering technique either kept or deliberately upgraded now that only one concept needs full development.

1. Do not resurrect the losing tiles as separate full-board generations "just in case"

— the grid already did the comparison work; if a second concept genuinely needs a fair look, run a fresh, smaller grid rather than promoting every candidate.

Workflow

1. Lock the palette and the rendering technique before drafting concepts

Decide these two things first, and hold them identical across every tile in the grid:

• Palette — the exact hex values in play (pull from an existing brand.md if one

exists, or from whatever the user states directly). A concept grid is not the place to explore palette — that's moodboard-generator's job. If no palette exists yet, either ask for one or run moodboard-generator first.

• Rendering technique — one named, concrete technique (e.g. "flat cut-paper

negative-space cutout," "single-line continuous contour," "matte 3D claymation render," "risograph two-tone halftone"). This is what keeps the grid reading as one coherent comparison sheet rather than six unrelated icon styles competing for attention — the reviewer should only be reacting to the idea in each tile, not to inconsistent rendering pulling their eye around.

2. Write each concept as one tight, abstract, single-mark sentence

Same discipline as logo-prompting's "honest specificity, not adjective soup" — each concept description needs a concrete visual referent an image model can actually render, not an adjective. Pick concepts genuinely distinct in form, not palette-swaps or minor pose variations of one idea; if two candidate concepts would produce near-identical silhouettes, that's one concept, not two, and one of them should be dropped or reworked before drafting the grid prompt. Aim for one clear noun or compact visual idea per concept — "a folded paper crane mid-unfold" not "something that evokes transformation and lightness."

3. Write the grid prompt

Structure, in order:

1. Frame it explicitly as an icon exploration grid / concept comparison sheet —

naming it this way (not "moodboard," not "icon set") is what keeps Ideogram rendering distinct, individually-labeled tiles instead of one blended composition.

1. State the grid dimensions (e.g. "a 2x3 grid of six square tiles," "a 3x2 grid of

six tiles") so the tile count and layout are unambiguous.

1. List each concept as its own captioned clause, in reading order, each naming its

own single-mark idea from step 2 plus a short label caption under the tile.

1. State the palette once, as the shared palette across every tile — never restate it

per-tile, since restating invites drift.

1. State the rendering technique once, as the shared technique across every tile, for

the same reason.

1. Apply the anti-slop gate (step 3.5 below) and fold its findings into the prompt.
2. Note clean, consistent tile composition — even tile spacing, a shared background

treatment behind every tile, thin dividing lines or a grid frame, and a consistent caption style/position under each tile — this is what keeps the sheet reading as one deliberate comparison tool rather than a collage.

Keep it one paragraph, same as the sibling board skills.

3.5. Run the anti-slop pre-generation gate

Reuse brand-identity-sheet's anti-slop gate exactly — see references/anti-slop-discipline.md in this skill, which is that file reused unmodified. Score the drafted prompt before calling generate_image. A concept grid has N times the surface area for a banned cliché that a single mark does — six tiles means six independent chances for the model to reach for a neural-network node, a glowing orb, a headset icon, or a gradient wash instead of the brand's actual motif vocabulary, so check each tile individually against the universal-ban list and the brand's own avoid-list, not just the prompt as a whole. Anything scoring under 3 on any axis for any tile means revise that tile's concept clause before generating.

4. Generate it

Call mcp__ideogram__generate_image with the paragraph prompt. Use style_type: "DESIGN" and aspect_ratio: "1x1" or "4x5" depending on tile count/layout — a grid of square tiles reads cleanest as a square canvas, a taller grid (more rows than columns) reads better as 4x5. Use rendering_speed: "QUALITY" — captions and tile dividers are text/composition-heavy and need to stay legible.

5. Save the artifact and route the winner forward

Save the prompt, the request_id, and the permalink to the project's branding folder (check for an existing logo-explorations/ or branding/ folder first and match it), clearly labeled as a Set-A / comparison artifact — e.g. concept-grid-.md — not as a finished deliverable. Record which concept was picked and why. Once a winner is named, hand it off as a separate follow-up generation to brand-identity-sheet or character-model-sheet per the two-stage handoff above — this skill's job ends at the pick, not at developing the winner further.

Reference files

• references/anti-slop-discipline.md — brand-identity-sheet's anti-slop gate,

reused unmodified, with a note on why a grid needs the per-tile check.

• examples/product-icon-concept-grid.md — a fully worked, fully fictional example:

a locked palette, a locked rendering technique, six distinct concepts, the full grid prompt, and a closing handoff note naming the winning concept.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: devkindhq
• Source: devkindhq/ideogram-ai-toolkit
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.