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
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
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
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
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
- Edit existing image (object removal, lighting swap, colorization, restoration, localization) → editing.md
- Character continuity across multiple images / panels → characters.md
- 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. Model default priors, banned booster words, capture pipeline instead of adjectives, located imperfections.
- Presentation slides → slides.md
- Sequential narrative (storyboard, comic, panel sequence) → storyboards.md
- Sketch → final, wireframes, structural input → structural.md
- 2D → 3D, floor plans, isometric → dimensional.md
- Vision analysis / image-to-prompt / style transfer from a reference image → vision-decomposer.md. 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. 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:
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
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