Concept

Design it first — concept sheet and storyboard

Drawing from nothing asks one model to invent a character, a pose, a camera, a palette and the placement of every pixel, all at once. It is the step where generated pixel art falls apart. Split the job:

• An image model owns the design — the look of the character, the key poses, the storyboard. That is what those models are good at.
• You own the pixel interpretation — reading that design into a canvas, a palette and a set of shapes, then drawing it with the rules.

The reference is a guide for proportion, pose and colour masses. It is never the pixels: a concept downscaled onto the canvas is anti-aliased mush, and no amount of cleanup turns it into pixel art (see What this must not do).

Two modes

Procedure

1. Start from the brief

aseprite:brief has already fixed canvas size, palette, view, light, outline and destination. Do not ask any of it again. The prompt needs colours, not a palette's name: for a preset, read its hex values from the knowledge://palettes resource (no document has to exist yet); for an open sprite, palette op="get".

2. Write the art spec

Short, concrete, in working notes. Everything the image model and, later, you will be held to:

• Subject and scenario — who or what, one line of personality, and for an animation what happens from the first frame to the last.
• Target — canvas W×H, how many pixels tall the subject stands, view and facing, light direction, outline style.
• Silhouette — heads tall, the two or three shapes that must read at 1× (a pointed hat, a long coat, an oversized sword).
• Palette — the exact hex list with roles: outline, skin ramp, cloth ramp, metal ramp, accent. Its size is the ceiling.
• Animation — cycle, frame count, one line per key pose, timing in ms, loop or one-shot. Plan it with rules://05-animation (or the animation-director agent) — the storyboard panels are these keys.
• Must not — text, background scenery, props that are not in the brief, camera changes between panels.

3. Build the prompt

In English whatever language the conversation is in — image models follow English prompts best. It has to produce something you can slice with reference op="import" grid=…, so the layout contract is not optional: equal panels, one row, same scale and ground line in every panel, flat background, no text.

For a character or prop — a concept sheet:

Pixel art character concept sheet of {subject}: {one-line personality}.
{view} view, {facing}, light from the {light}. Low-resolution pixel art look:
the character is about {height}px tall on a {W}×{H} sprite, hard pixel edges,
no anti-aliasing, no gradients, {outline style} outline.
Palette — use only these colours: {hex list with roles}.
Layout: {n} panels in a single horizontal row, equal width, evenly spaced,
the same character at the same scale in every panel, feet on the same ground
line: {panel 1: front | panel 2: side | panel 3: back | …}.
Flat plain {#ffffff or #808080} background, no scenery, no text, no labels,
no borders, no watermark.

For an animation — a storyboard, one panel per frame:

Pixel art animation storyboard of {subject} performing a {cycle}.
{view} view, facing {facing}, camera locked, light from the {light}.
Low-resolution pixel art look: the character is about {height}px tall on a
{W}×{H} sprite, hard pixel edges, no anti-aliasing, {outline style} outline.
Palette — use only these colours: {hex list with roles}.
Layout: exactly {N} panels in a single horizontal row, equal width, evenly
spaced, identical character size and ground line in every panel — each panel
is one frame, left to right:
1. {key pose 1}
2. {key pose 2}
…
{N}. {key pose N}
Flat plain {background} background, no motion lines, no text, no numbers,
no borders, no watermark.

Ask for both when the request is an animated character: the sheet fixes the design, the storyboard fixes the motion. Fill every {…} — a prompt with a placeholder left in it produces a generic character.

4. Offer the choice — once, in one message

Put the prompt in a single fenced code block, so it copies in one action. If the brief is still awaiting confirmation, this is the same message — decisions, prompt, choice — not a second round trip. Then offer exactly these options, through the harness's question tool when it has one (Claude Code's AskUserQuestion, omp's ask), otherwise as a numbered list:

1. "I'll generate references" — the user pastes the prompt into any image model (ChatGPT, Gemini, Midjourney, …) and sends the result back. Ask for the image as a file path — saved to disk, or dragged into the chat where the harness turns that into a path — because reference import reads files, not pixels you have only seen in the conversation.
2. "Continue without references" — go straight to drawing from the spec. The spec still binds: it is the design now.

If your harness has an image-generation tool, add it as a third option ("you generate it"), never as a silent default. When the user chooses option 1, stop and wait for the image; do not start drawing in the meantime.