Update to Simplified

{
  "name": "generate-image",
  "description": "Generate AI images with Simplified — text-to-image, image editing, and reference-guided generation across Flux, Google (Gemini/Imagen), OpenAI GPT Image, Ideogram, Stable Diffusion, Qwen and Seedream. Use when the user asks to create, generate, make, draw, or design an image, photo, picture, graphic, logo, poster, banner, icon, or illustration from a description.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 754
    },
    {
      "relative_path": "assets/simplified-icon-256.png",
      "size_in_bytes": 8191
    },
    {
      "relative_path": "assets/simplified-logo.png",
      "size_in_bytes": 6726
    },
    {
      "relative_path": "assets/simplified-treadmark-logo.png",
      "size_in_bytes": 18491
    }
  ],
  "skill_md_contents": "---\nname: generate-image\ndescription: >-\n  Generate AI images with Simplified — text-to-image, image editing, and\n  reference-guided generation across Flux, Google (Gemini/Imagen), OpenAI GPT\n  Image, Ideogram, Stable Diffusion, Qwen and Seedream. Use when the user asks to\n  create, generate, make, draw, or design an image, photo, picture, graphic,\n  logo, poster, banner, icon, or illustration from a description.\n---\n\n# Generate AI Image\n\nGenerate an image from a text prompt using Simplified, across many leading AI\nproviders, and return a viewable image URL (plus an asset id you can reuse).\n\n## What it can do\n\n- **Text-to-image** (`capability: \"prompt\"`) — make an image from a description.\n- **Image editing / image-to-image** (`capability: \"reference_image\"`) — transform\n  or edit using one reference image.\n- **Multi-reference composition** (`capability: \"multiple_images\"`) — guide with\n  several reference images (supported on some models).\n\nGood for: product shots, hero/banner images, social graphics, illustrations,\n3D-style renders, icons/logo concepts, photoreal scenes, and **text rendered inside\nthe image** (posters, quote cards, ad headlines).\n\n## How to use it\n\n1. **Discover** — call `api_getModelFields(type: \"image\")` to get the current list of\n   models, capabilities, and credit costs. Filter out models that cannot satisfy the\n   requested capability; don't choose on model name alone.\n2. **Choose the model** — use [Model selection](#model-selection). For the selected\n   model, call `api_getModelFields(type: \"image\", model_id, capability)` and use its\n   exact `parameters` schema. This is the source of truth; don't guess model ids,\n   capabilities, field names, or costs.\n3. **Choose storage** — `transient` for a one-off, `asset` to reuse the image (e.g.\n   post it via the `simplified-social` skill).\n4. **Explain the choice when it matters** — before a costly or ambiguous request,\n   name the selected model, why it fits, and the discovered credit cost. If the user\n   explicitly chose a model, honor it when it supports the requested capability.\n5. **Generate** — call `api_generateImage` with `parameters` matching the discovered\n   schema. This spends credits.\n6. **Present the result** — show the returned URL as a link, never embedded (see\n   [Presenting the result](#presenting-the-result)).\n\nFor an ordinary prompt-only request, use the quality-first default below after\nconfirming it is still available. Always inspect live fields for reference-image,\nmulti-image, exact-size, quality, or resolution requests.\n\n## The request\n\n### The tools\n\n- **`api_getModelFields`** — discover available models and the per-(model, capability)\n  field schema. Read-only, spends **no credits**. Call it first.\n- **`api_generateImage`** — **consumes paid AI credits**.\n\n### Fields\n\nTop-level fields for `api_generateImage`:\n\n- `model` — a model id from `api_getModelFields` (e.g. `google.gemini-3.1-flash-image-preview`).\n- `capability` — `prompt` | `reference_image` | `multiple_images`.\n- `storage` — see [Storage](#storage) (default `transient`).\n- `parameters` — a **required nested object**; never flatten its fields to the top\n  level, and put the prompt text in `parameters.prompt` (not in `capability`).\n\nThe **exact keys inside `parameters` vary by model** — get them from\n`api_getModelFields(type: \"image\", model_id, capability)`, don't assume. They differ\nin real ways: most models take `aspect_ratio`, but OpenAI GPT Image uses `size` +\n`quality` + `count`, Gemini adds `image_size`, Flux 2 uses `resolution`, and the\nreference-image field is variously named `input_image`, `image_prompt`,\n`reference_images`, `source_image`, or `style_reference_images`.\n\n### Resolving Simplified asset references\n\nTreat a Simplified `asset_id` as the canonical reference, but follow the live model\nschema at the generation boundary. When a model field is a URL or URL list (for\nexample Gemini `reference_images`):\n\n1. Call `api_getAsset` with the permanent asset UUID.\n2. Require `status: 4` (`DONE`) and the expected `asset_type` before generating.\n3. Pass the current `file_url` returned by `api_getAsset` into the model-specific\n   reference field. If the URL is signed, preserve its complete query string and use\n   it before expiry.\n4. Do not trust a cached URL copied from a brand-kit record when an `asset_id` is\n   available. Brand records can contain stale or malformed derived URLs; resolve the\n   ID immediately before generation instead.\n\nIn short: **IDs at rest, URLs at the model boundary, IDs downstream.** Do not pass a\nclient-local path to the hosted connector.\n\n### Storage\n\n| `storage` | Behavior |\n|---|---|\n| `transient` | **Default.** Temporary URL, not saved, expires. Best for one-off images. |\n| `asset` | Persistent — no expiry, returns an `asset_id`. Use when you want to **reuse** the image, e.g. attach it to a post via the `simplified-social` skill (pass the `asset_id` in `media`). |\n| `default` | Saved to your AiImageArt gallery. |\n\n### Examples\n\n**Text-to-image (default, transient):**\n```json\n{ \"model\": \"google.gemini-3.1-flash-image-preview\", \"capability\": \"prompt\", \"storage\": \"transient\",\n  \"parameters\": { \"prompt\": \"A white ceramic coffee cup on a clean white background\", \"aspect_ratio\": \"1:1\" } }\n```\n\n**Keep it to reuse / post to social (asset):**\n```json\n{ \"model\": \"google.gemini-3.1-flash-image-preview\", \"capability\": \"prompt\", \"storage\": \"asset\",\n  \"parameters\": { \"prompt\": \"product hero shot of sneakers\", \"aspect_ratio\": \"4:5\" } }\n```\n\n**Edit / reference-guided** — the reference field name is model-specific; take it from\n`api_getModelFields` (here `input_image` for a Flux Kontext model, not a guessed name):\n```json\n{ \"model\": \"flux.flux-kontext-pro\", \"capability\": \"reference_image\", \"storage\": \"asset\",\n  \"parameters\": { \"prompt\": \"put this logo on a t-shirt\", \"input_image\": \"<asset_uuid_or_https_url>\" } }\n```\n\n## Model selection\n\nChoose for the requested outcome, not provider popularity. These routes are maintained\ndefaults, but model availability, capabilities, parameters, and credits can change;\n`api_getModelFields(type: \"image\")` remains authoritative.\n\n| User need | Preferred model | Why / tradeoff |\n|---|---|---|\n| Normal social image, product shot, illustration, character continuity, or general edit | `google.gemini-3.1-flash-image-preview` | **Quality-first default.** Strong all-around prompt following and reference fidelity. Do not interpret “Flash” as the cheapest option. |\n| Complex professional design, dense typography/layout, menu, invitation, high-fidelity product mockup, factual visualization, or explicit 4K | `google.gemini-3-pro-image-preview` | Premium quality and instruction handling; slower and typically costs more. Use only when the request benefits from it. |\n| Budget-sensitive generation or explicit GPT Image request | `openai.imgen-2` | The catalog's `credits_per_image` is a **baseline**, not the final charge. Cost varies with `size`, `quality`, and `count`. Use the live API field `quality: \"auto\"` (the operational “effort auto” setting) unless the user requests a different quality. It uses `size` rather than `aspect_ratio`. |\n| Short headline or typography-first poster/banner | `ideogram.ideogram-v3-turbo` | Specialized text rendering. Prefer Gemini Pro when the design also requires a dense or complex professional layout. |\n| Targeted edit with a single source image | `google.gemini-3.1-flash-image-preview`; `flux.flux-kontext-pro` when explicitly requested or better suited by live metadata | Default to Gemini for fidelity. Flux Kontext is a specialized alternative; inspect its `input_image` contract first. |\n| Many reference images or exact reference limits | Best compatible model returned live | Filter by `multiple_images` and the discovered reference limit. Never assume every model accepts the same number or field name. |\n| User names Flux, Seedream, Qwen, Stable Diffusion, or another available model | The requested model, if compatible | Respect an explicit preference. Otherwise do not automatically route to an unvalidated specialist merely because it is available or cheaper. |\n\n### Routing rules\n\n1. Infer the hard constraints: capability, reference count, aspect ratio/size,\n   resolution, text/layout complexity, budget, and any explicit provider choice.\n2. Filter the live catalog by those constraints.\n3. Use Gemini 3.1 Flash when no stronger constraint applies. Upgrade to Gemini 3 Pro\n   only for the professional-design cases above. Consider GPT Image 2 when minimizing\n   credits is explicit or as the first fallback, but state that its live catalog rate\n   is only a baseline and the final charge varies with `size`, `quality`, and `count`.\n   Default to `size: \"auto\"`, `quality: \"auto\"`, and `count: 1` unless the request\n   requires different values.\n4. For a typography-first graphic, choose Ideogram Turbo; for dense layout or 4K,\n   choose Gemini Pro instead.\n5. Never silently change models after an error. Report the failure and proposed\n   fallback with its live credit cost, then regenerate only when the user's existing\n   intent clearly authorizes the additional spend.\n\nWhen a request is ambiguous and the choice materially changes cost or output, offer\nthe most relevant two choices, leading with the recommended model. Do not dump the\nentire catalog on the user.\n\n## Response\n\nThe response shape depends on `storage`:\n\n- **`transient` (default)** — `result` is a list of **URL strings**:\n  ```json\n  { \"status\": \"SUCCESS\", \"detail\": { \"result\": [\"https://replicate.delivery/…/out-0.webp\"], \"transient\": true } }\n  ```\n  Read `detail.result[0]` (a URL string). No `asset_id` — the URL is temporary.\n\n- **`asset`** — `result` is a list of **objects** with a reusable id:\n  ```json\n  { \"status\": \"SUCCESS\", \"detail\": { \"result\": [{ \"asset_id\": \"<uuid>\", \"url\": \"https://…/image.webp?Expires=…\" }], \"transient\": false, \"storage\": \"asset\" } }\n  ```\n  Read `detail.result[0].url` (the image; **signed URL — expires**) and\n  `detail.result[0].asset_id` (permanent — hand off to `simplified-social`'s `media`).\n\nOutput format varies by model and provider. Inspect the returned asset or response\nmetadata instead of assuming WebP; for example, Gemini may return JPEG.\n\n## Presenting the result\n\n**Never embed the returned image URL with Markdown image syntax** (`![](url)`), and\nnever do anything that makes the client fetch/render the image inline. Always present\nthe result as a **plain URL or a Markdown link** the user can click:\n\n- ✅ `Here's your image: https://…/out-0.webp`\n- ✅ `[View generated image](https://…/out-0.webp)`\n- ❌ `![generated image](https://…/out-0.webp)`\n\nReasons: these URLs are signed and **expire**, inline rendering fails or shows a\nbroken image, and clients like Codex otherwise try to display the asset instead of\nhanding the user a usable link — poor UX. When `storage:\"asset\"`, also surface the\npermanent `asset_id` (as text) so it can be reused with `simplified-social`.\n\n## Gotchas\n\n- **Discover before generating.** Call `api_getModelFields` to confirm the model id\n  and `parameters` schema — it eliminates 400 errors on invalid/missing keys and\n  prevents routing from stale model or credit assumptions.\n- **Resolve asset-backed references before generating.** Use `api_getAsset`, require\n  `status: 4`, and pass its current `file_url` when the live model field expects a\n  URL. Keep the source `asset_id` for future runs.\n- **Generation spends credits.** If the request is ambiguous, restate what you'll\n  generate and confirm once. If it's explicit, proceed.\n- **Do not overstate GPT Image 2 pricing.** Treat `credits_per_image` as baseline\n  metadata. Final usage varies with `size`, `quality`, and `count`. The live API calls\n  the effort control `quality`; use `quality: \"auto\"` for the usual “effort auto”\n  behavior and never describe the baseline as the guaranteed charge.\n- `429` = AI credits exhausted; tell the user plainly and don't retry.\n- On error, report it; don't silently retry.\n\n## Example prompts to try\n\n- \"A minimalist product photo of a white ceramic coffee cup on a clean white background, soft studio lighting\"\n- \"A vibrant 3D render of a friendly robot mascot, pastel colors, studio lighting, 1:1\"\n- \"A cinematic 16:9 landscape of snowy mountains at golden hour\"\n- \"A flat vector app icon of a paper plane, rounded corners, blue gradient\"\n- \"A bold quote card that says 'Ship it' in modern type\" (use `ideogram.ideogram-v3-turbo` for crisp text)\n"
}