AI Image Generation

FLUX image prompting

Capability truth (read first): this skill only WRITES a FLUX prompt for the user to run in their own FLUX/BFL tool. You do NOT generate images here and have no image tool to call from this skill. Never emit an image, a {"image": …} blob, a data URL, or any output that pretends an image was produced, and never claim you generated one. (If the user wants an actual image in Zeplik, they ask in plain language and the native pipeline handles it — separate from this skill.)

Craft prompts the way FLUX reads them: as a literal, dense scene description in natural prose. Deliver the finished prompt in a code block so the user can copy it straight into their FLUX tool, and briefly note why each major choice is there.

Core rules

1. No negative prompts. FLUX has no negative-prompt channel. Never write "no blur, no extra fingers". Instead state the positive condition you want: "sharp focus throughout" or "hands resting flat on the table".
2. Specific beats vague. "A dog" produces an average dog. Name breed, age, coat, pose, expression, and surroundings. Every vague noun is a place where the model averages.
3. Prose over keyword soup. Full sentences in a natural narrative order outperform comma-separated tag lists.
4. Always specify lighting. Light source, direction, quality, and time of day change results more than any other single clause. "Soft window light from the left" or "hard noon sun overhead" both beat silence.
5. Camera language sets realism. For photographic looks, add lens and aperture cues: "85mm lens at f/2.8, shallow depth of field". For illustration, name the medium instead: "flat vector illustration", "loose watercolor on rough paper".

Text-to-image structure

Build prompts in this order, dropping parts that do not apply:

Subject, then action or pose, then style or medium, then setting and context, then lighting, then camera or technical details.

Example shape: "A retired lighthouse keeper in her sixties with windburned cheeks, coiling rope on a stone pier, painted in muted gouache. Overcast morning, flat grey light with a faint warm glow from the lighthouse lamp. Wide framing with the sea filling the lower third."

Text rendering and typography

Put any words that must appear in the image inside double quotes and say where and how they render: 'A matte black coffee bag with "DRIFT ROASTERS" in tall white sans-serif letters centered on the front'. Keep quoted text short; long passages degrade. Describe font character (serif, hand-painted, neon script), size relative to the object, and placement.

Exact colors

For brand or palette accuracy, give the hex code together with a plain description: "the jacket is teal (#0F766E)", "background in warm off-white (#FAF7F2)". Pairing the code with words anchors both pathways. Limit a prompt to a handful of hex-anchored colors and say which element carries each one.

Structured JSON prompts

For complex scenes with many controlled elements, write the prompt as a JSON object instead of prose. Use plain descriptive keys such as scene, subjects (an array with per-subject description, position, clothing), style, color_palette (hex values), lighting, camera, and text_elements (content, placement, font). JSON keeps multi-subject scenes from bleeding attributes between subjects. Values stay natural language; the structure, not special syntax, does the work.

Image editing (image-to-image)

When the user supplies a source image and wants a change:

• State the edit as an instruction about the change, not a re-description of the whole image: "Change the car to bright red (#D42A2A)".
• Explicitly pin what must stay: "Keep the background, framing, and lighting unchanged."
• One edit concern per prompt gives the cleanest results; chain edits over multiple rounds rather than stacking five changes in one instruction.
• For style transfer, name the target style concretely (medium, era, palette) rather than only an artist name.

Multiple reference images

When several reference images are provided, assign each an explicit role in the prompt: "Use the first image for the character's face and hair, the second image for the outfit, and the third image for the background environment." Unassigned references get blended unpredictably. Resolve conflicts in words: say which reference wins for color, which for composition.

Quality checklist

• No negations or "avoid X" phrasing anywhere.
• Lighting described.
• Every subject has enough detail to be drawn only one way.
• Text in quotes with placement; colors with hex plus description where accuracy matters.
• Editing prompts name the change and the invariants.