/generating-images

name: generating-images
description: ALWAYS read this skill before generating or editing any image, or calling image_generate — text-to-image or image-to-image, simple or complex. Turns a text brief, optionally guided by reference images (a character, a style to follow, an existing photo to edit), into a still image. Analyzes intent, routes to the best model, and structures the prompt. Use whenever the user asks to generate, create, make, design, or edit an image — except commercial photography of a product being sold, which belongs to generating-product-photos.
license: Apache-2.0
metadata:
  version: "0.1.0"
  category: creative
  summary: "Routes your brief to the best of 10+ SOTA image models (like Nano Banana, GPT Image), writes detailed prompts as per model specifications and generates multiple variations instantly."

Image Generation

Turn a brief — text, optionally with reference image(s) — into a still image via the `image_generate` tool. Two decisions drive quality: **which model** (always) and **which format recipe** (only when the deliverable is a known one).

**Before you route:** if the brief is commercial **product photography** — a packshot, a product in a styled scene, a hero or banner for a product, an on-model or try-on shot, or restyling an existing product photo — hand off to the **`generating-product-photos`** skill, which owns that surface. If it is a **product advertisement** — a headline, an offer or a call to action drawn over the product, a promotional before/after or comparison — hand off to **`generating-image-ads`**. Stay here for everything else: general images, graphics and posters, portraits, illustrations, cinematic stills, infographics, and one-off reference edits.

Workflow

Step 1: Route to a model

Route by what the image has to **do** — read the brief for intent. The descriptions below are signals to weigh, not a literal router — the examples are illustrative, not a checklist to match against. Then **read the chosen model's prompt guide before writing anything**.

| Route when the brief is about… | Model | Prompt guide | | --- | --- | --- | | Rendering **words** legibly, or a designed layout where elements sit in deliberate positions — for example a poster, ad, banner, thumbnail, or infographic | `gpt-image-2` | `references/prompt-gpt-image-2.md` | | A **drawn or rendered** look rather than a photograph — for example a cartoon, anime, illustration, flat vector, or 3D render | `nano-banana-2` | `references/prompt-nano-banana.md` | | A convincing **real person**, or a photographic frame with deliberate cinematography — for example a creator or influencer portrait, UGC, or a film-like still | `nano-banana-pro` | `references/prompt-nano-banana.md` | | **Altering a supplied image** — for example swapping a background, removing or replacing an element, or restaging the scene | `gpt-image-2` | `references/prompt-gpt-image-2.md` | | Altering a supplied image where a **real face must stay recognisable** | `seedream-5` | `references/prompt-seedream.md` |

When more than one applies, take them in this order:

1. The user named a model → use it. 2. Legible text is required → `gpt-image-2`, even over a person or a scene. 3. The look is non-photographic → `nano-banana-2`, even when a person is involved. 4. A supplied image is being altered → `gpt-image-2`; switch to `seedream-5` when a real person must stay recognisable. 5. A supplied image is only a **style or mood cue** (match this look, don't edit that image) → `nano-banana-2`, or `seedream-5` if a specific face must carry over too.

If nothing clearly fits — an ordinary object or scene — use `nano-banana-2` (read `references/prompt-nano-banana.md`); or call `list_image_models` and pick by `strengths`. Use `grok-imagine` (`references/prompt-grok.md`), `flux-2-pro`, or `flux-2-klein-4b` (both `references/prompt-flux.md`) only when the user names them.

Step 2: Read a format recipe if one fits (optional)

Many briefs are a standard deliverable with a layout worth following. Check the table below. **If a row matches, read that recipe and follow its sections and example. If none matches, skip this step** — the model's prompt guide is enough.

| Deliverable | Recipe | | ------------------------------------ | -------------------------------- | | Poster / ad / banner / social graphic | `references/format-poster.md` | | Portrait / avatar / influencer | `references/format-portrait.md` | | Cinematic still | `references/format-cinematic.md` |

Step 3: Get any missing inputs

Ask only for what would change the result and can't be sensibly defaulted — for example the exact text that has to appear on the image, a source image for an edit, or the desired aspect ratio when that decides the crop. Bundle everything you need into one ask rather than a back-and-forth; if the user has signalled they don't want questions, choose sensible defaults, state them in a line, and proceed.

Step 4: Look at the reference images — only when references are supplied

Call `image_analysis` on each reference, asking whatever the prompt will have to carry — enough to name what must stay fixed and to avoid contradicting the source.

Skip when you already know what's in the image — you generated it this turn, or it arrived with a description.

If what you find changes the routing — a face to keep recognisable, a stylized source — go back to Step 1.

Step 5: Write the prompt

Build the prompt as the chosen model's prompt guide specifies. If you read a format recipe, start from its sections and worked example. With a reference image, use what Step 4 turned up to say what stays fixed (the product, the face, the palette) and what changes.

**Write a complete, self-contained prompt** — the model sees only this one prompt. Specify the whole frame: subject, wardrobe / materials, setting, lighting, colour palette, camera / lens, medium