Image

Image Prompting — Nano Banana & GPT Image 2.5

This skill writes image prompts. It does not generate images. The output is: model name + quality / size / aspect ratio + the prompt itself.

The body of this SKILL.md is intentionally thin so you cannot fake a result by reading it alone. The actual rules — what the models reward, what they punish, how to phrase a 5-slot template, when to add quality: high, when to use image grounding — live only in the reference files.

Route first — is this actually an image-prompt task?

• Motion, clips, montage (Seedance, Kling, Veo, any image-to-video): use the sibling video skill. This skill's storyboard and keyframe outputs feed it.
• No idea or script yet (user wants a concept or an ad scenario, not a picture): if the creative-director skill is installed, start there — it develops ideas and scripts for commercials and beyond (github.com/smixs/creative-director-skill).
• A concrete image is needed — this skill. Continue below.

Mandatory reading order — DO NOT WRITE A PROMPT WITHOUT THIS

Past attempts to write prompts directly from this skill body produced lazy, generic results. Each model has its own physics; common rules collapse into mush when applied without model-specific syntax. Read in this order before producing any prompt:

Step 1 — always read first → models.md [blocked]

Decide: Nano Banana (NB2 or NBP) or GPT Image 2.5 (Flare for speed, Sunburst for precision edits). The choice changes the prompt syntax fundamentally — natural-language paragraphs vs. labeled 5-slot template, quality settings, which features exist (image grounding only on NB, EXACT TEXT discipline only on GPT Image, etc.).

If the user named a model — confirm and proceed. If not — pick using the table in models.md, then state your choice in the output header.

Step 2 — read one model file (the one you picked)

• Nano Banana → nano-banana.md [blocked] Image grounding for real locations. Extreme aspect ratios (1:8, 8:1, 4:1). Thinking mode. JSON for 5+ elements. Up to 14 reference images. Why you must NOT write 50mm / f-stop / ISO numbers.

• GPT Image 2.5 → gpt-image.md [blocked] 5-slot template (Scene / Subject / Important Details / Use Case / Constraints). Anti-slop banned-words list. quality: low / medium / high / xhigh / max as a deliberate fidelity lever. Size constraints (multiples of 16, max 3:1, up to 4K 3840×2160). Two-column edit logic (Change / Preserve / Constraints). Up to 16 reference images with explicit roles.

The model file is non-negotiable. Skipping it is the single biggest cause of weak prompts.

Step 3 — always read after the model file → golden-rules.md [blocked]

Universal rules that apply to both models: start with a verb, positive framing, hex colors, quote text, edit don't re-roll, one change per iteration, reference images.

Step 4 — task-shaped reading (load only what matches the request)

Pick zero or more, depending on what the user asked for:

• Text in image, infographic, diagram, multilingual rendering → text-rendering.md [blocked]
• Edit existing image (object removal, lighting swap, colorization, restoration, localization) → editing.md [blocked]
• Character continuity across multiple images / panels → characters.md [blocked]
• The image must pass as a real photograph (portrait, reportage, UGC, casting, product-in-hand) — or the user says the result "looks AI", "too glossy", "not like the reference" → de-slop.md [blocked]. Model default priors, banned booster words, capture pipeline instead of adjectives, located imperfections.
• Presentation slides → slides.md [blocked]
• Sequential narrative (storyboard, comic, panel sequence) → storyboards.md [blocked]
• Sketch → final, wireframes, structural input → structural.md [blocked]
• 2D → 3D, floor plans, isometric → dimensional.md [blocked]
• Vision analysis / image-to-prompt / style transfer from a reference image → vision-decomposer.md [blocked]. Load this whenever the user attaches an image and asks to recreate, match, decompose, or transfer its style.
• Multi-panel compositions (grids, collages, storyboard sheets in ONE image) → multi-panel.md [blocked]. 9-cell TVC grids, 2x2 portrait grids, 3-panel campaign collages, 4x3 borderless grids, 6-frame cinematic sequences, before/after splits, 12-panel storyboard posters.
• Industry pattern libraries — proven prompt templates by vertical. Load the matching file:
  • E-commerce product shots → patterns/ecommerce.md [blocked]
  • Fashion editorial campaigns → patterns/fashion-editorial.md [blocked]
  • Food & beverage advertising → patterns/food-beverage.md [blocked]
  • Cinematic portraits → patterns/portrait-cinema.md [blocked]
  • Posters & illustration → patterns/poster-illustration.md [blocked]. Also holds the Style DNA + Reject Checklist method — how to fix a strong visual style in four lines before generating and how to judge the returned image against it. Usable with any pattern in this list, not only posters.
  • Character design (turnarounds, expression sheets, outfit grids) → patterns/character-design.md [blocked]
  • UI mockups & social media formats → patterns/ui-social.md [blocked]

Step 5 — read for production language → creative-direction.md [blocked]

Studio-quality vocabulary for lighting design, camera and hardware, color grading and film stock, materiality and texture. Read when you need precise terms beyond what golden-rules.md covers.

Step 6 — read if structuring a complex prompt → prompt-framework.md [blocked]

Universal element checklist (subject, context, action, environment, camera, lighting, mood, materials, palette, format), detail modes (concise / standard / verbose / cinematic verbose), parameterized templates, output structure with parameters and exclusions.

Output format

When you return the prompt, structure it like this:

Model: <nano-banana-2 | nano-banana-pro | gpt-image-2.5-flare | gpt-image-2.5-sunburst>Quality: <low | medium | high | xhigh | max>   (only for gpt-image-2.5)Size / Ratio: <e.g. 1536×1024 or 16:9>
Prompt:<the prompt text, ready to copy>
Notes:- <anything you inferred or assumed because the user did not specify>

For edits, also include an explicit preserve-list (mandatory for gpt-image-2.5, recommended for nano-banana):

Change: <one concrete thing>Preserve: <face, pose, lighting, framing, geometry, ...>Constraints: <no extra objects, no drift, ...>

Final response style

Prefer: ready-to-copy prompts, hex colors, concrete materials, named compositions, model-specific syntax (5-slot for GPT Image, natural prose for Nano Banana).

Avoid: tag soup ("cool, modern, 4k"), vague praise ("stunning, epic, masterpiece" — actively hurts GPT Image 2.5), negative framing ("no people, no cars" — invert to positive), external comparisons ("like Apple ad" — describe the visual properties instead), numerical lens parameters in Nano Banana prompts (it ignores them).

Author: Serge Shima (t.me/aimastersme · sergeshima.com · aimasters.me) · License: CC BY 4.0 — attribution required · Source: smixs/visual-skills