← Sceneform AI Ads & InfluencersCONTENT HISTORY
WHAT CHANGED
Update to Sceneform AI Ads & Influencers
Snapshot Nov 16, 2026 · 18:17 UTC · version 0.9.0
Collection source: downloaded plugin package.
First saved snapshot
No earlier snapshot is available to establish a change.
Save Sceneform AI Ads & Influencers and check future changes in Following.
Up to 3 plugins free · no card required.
Compare saved observations
Download comparison JSONFull technical diff · 0 changed fields
Full snapshot data
{
"description": "Generate images, video, speech, music, sound effects, voice changes, lip sync, and talking avatars with Sceneform, and edit, animate, upscale, or reframe existing media. Use when the user wants any of these from a brief, a reference, or a file they already have, including a viral remix, a narrated storybook explainer, or a restyled clip. NOT for ads or carousels (sceneform-ad-studio), BrandKits or Personas (sceneform-brand), animated text and graphics (sceneform-motion), or posting (sceneform-publish).",
"included_files": [
{
"relative_path": "references/audio.md",
"size_in_bytes": 4955
},
{
"relative_path": "references/cli.md",
"size_in_bytes": 2168
},
{
"relative_path": "references/image.md",
"size_in_bytes": 11260
},
{
"relative_path": "references/models.md",
"size_in_bytes": 21204
},
{
"relative_path": "references/prompt-flux-image.md",
"size_in_bytes": 4699
},
{
"relative_path": "references/prompt-flux-video.md",
"size_in_bytes": 4567
},
{
"relative_path": "references/prompt-gpt-image.md",
"size_in_bytes": 3840
},
{
"relative_path": "references/prompt-hailuo.md",
"size_in_bytes": 4714
},
{
"relative_path": "references/prompt-ideogram.md",
"size_in_bytes": 3948
},
{
"relative_path": "references/prompt-kling-video.md",
"size_in_bytes": 3338
},
{
"relative_path": "references/prompt-ltx-video.md",
"size_in_bytes": 3957
},
{
"relative_path": "references/prompt-music.md",
"size_in_bytes": 849
},
{
"relative_path": "references/prompt-nano-banana.md",
"size_in_bytes": 4123
},
{
"relative_path": "references/prompt-other-models.md",
"size_in_bytes": 4205
},
{
"relative_path": "references/prompt-pixverse.md",
"size_in_bytes": 3373
},
{
"relative_path": "references/prompt-qwen.md",
"size_in_bytes": 3552
},
{
"relative_path": "references/prompt-recraft.md",
"size_in_bytes": 3331
},
{
"relative_path": "references/prompt-seedance.md",
"size_in_bytes": 4375
},
{
"relative_path": "references/prompt-seedream.md",
"size_in_bytes": 4303
},
{
"relative_path": "references/prompt-soul.md",
"size_in_bytes": 2914
},
{
"relative_path": "references/prompt-speech.md",
"size_in_bytes": 2979
},
{
"relative_path": "references/prompt-veo.md",
"size_in_bytes": 3634
},
{
"relative_path": "references/prompt-vidu.md",
"size_in_bytes": 3220
},
{
"relative_path": "references/prompt-wan-image.md",
"size_in_bytes": 2417
},
{
"relative_path": "references/prompt-wan-video.md",
"size_in_bytes": 4007
},
{
"relative_path": "references/short-videos.md",
"size_in_bytes": 2748
},
{
"relative_path": "references/video.md",
"size_in_bytes": 8010
}
],
"name": "sceneform-generate",
"skill_md_contents": "---\nname: sceneform-generate\ndescription: Generate images, video, speech, music, sound effects, voice changes, lip sync, and talking avatars with Sceneform, and edit, animate, upscale, or reframe existing media. Use when the user wants any of these from a brief, a reference, or a file they already have, including a viral remix, a narrated storybook explainer, or a restyled clip. NOT for ads or carousels (sceneform-ad-studio), BrandKits or Personas (sceneform-brand), animated text and graphics (sceneform-motion), or posting (sceneform-publish).\nlicense: MIT\nmetadata:\n author: sceneform\n version: \"0.9.0\"\n---\n\n# Generate with Sceneform\n\nYou are the art director and the prompt writer. Sceneform sends your prompt\nto the model exactly as you write it, so the result is only as specific as\nthe prompt. The craft lives in reference files: read one by calling\n`load_skill` with this skill's name and the file's `path`, or open it when\nthis skill is installed locally.\n\n## Steps\n\n### 1. Read the brief\n\nName the kind (image, video, speech, music, and so on), the deliverable,\nwhere it will appear, and how many. Take every value the user gave. Fill the\nrest with defaults: one output, the aspect ratio of the destination (the\ncraft references list them), and the style the subject and brand imply.\n\nAsk only about a gap that blocks a useful result, and bundle up to three\nshort questions in one message:\n\n- the subject, when nothing describes it;\n- whose face, when a specific real person must appear and there is no photo\n or Persona;\n- what to change, when the user rejects a result without saying why.\n\nModel, lens, lighting, and prompt wording are your decisions. When the user\nsays \"just make it\", \"go ahead\", or similar, ask nothing: state the plan in\none sentence and generate.\n\nWhen the work is for the user's brand or product (\"our\", \"my shop\", a\nproduct name), read it first with `list_brand_kits` and use its facts.\nClaims, prices, colors, and label text come only from the user or the kit.\nA UGC or creator video, or any ad for a brand or product, belongs to\nsceneform-ad-studio: `generate_video_ad` writes the script and shots from\nyour concept, which beats a hand-written video prompt.\n\nDone when every value has a source (the user said it, or you chose the\ndefault) and you can state the plan in one sentence.\n\n### 2. Load the craft for the kind\n\n| The request | Read |\n| --- | --- |\n| An image, an image edit, upscale, background removal, reframe | [references/image.md](references/image.md) |\n| A video clip, an animated image, a Reshoot of a clip | [references/video.md](references/video.md) |\n| Speech, music, sound effects, voice change, lip sync, avatar | [references/audio.md](references/audio.md) |\n| A viral remix, a storybook explainer, a restyled clip | [references/short-videos.md](references/short-videos.md) |\n| Working from the `sceneform` CLI | [references/cli.md](references/cli.md) |\n\nThen pick the model by the job. A model the user names wins; otherwise start\nhere:\n\n| The job | Model |\n| --- | --- |\n| An ad, a UGC or creator video, a product demo | Not `generate_video`: sceneform-ad-studio, `generate_video_ad` |\n| Any other video: people talking, dialogue and sound, several shots, references | Seedance 2.5 (the default) |\n| Trying a video idea cheaply | Seedance 2.5 at `resolution: \"draft\"`, then `finish_draft` |\n| Fast physical action: sports, dance, stunts | Kling 4.0 |\n| A clip that must follow reference clips closely, or 4K | Gemini Omni Flash 1.1 |\n| Many references at once, or copying a reference clip's motion | H3 Max |\n| An image with text, a design, an ad, a product photo, or an edit | GPT Image 2.6 Sunburst (the default) |\n| A photoreal scene, an illustration or cartoon character, or many reference images combined | Nano Banana Pro |\n| An edit of a real photo that must keep the person's face | Seedream 5.0 Pro |\n| A candid phone photo of a person, with no product or text in frame | Soul 2.0 |\n\nFor a job the table does not cover, [references/models.md](references/models.md)\nlists every model by kind with what it is best and worst at, what it takes,\nand what it makes.\n\nFor an image or a video, read the chosen model's own guide. Call\n`query_generation_options` with `action: \"search\"`, the `kind`, the model's\nname as `query`, and `available_input_kinds` for any references (for a job\nthe table does not cover, `action: \"recommend\"` with a few task words\ninstead). Pick the first option whose `input_roles` cover every reference\nyou will pass; when the model cannot take them, say so and offer the closest\none that can. Then call it with `action: \"get\"` and that `option_ref`. Its\n`prompt_guide` says how that model reads a prompt, how long a good one runs,\nand whether it renders sound. Pass the same `option_ref` to `generate_image` or `generate_video` so\nthe guide and the model match. When the tool is out of reach, the model's\nrow in `references/models.md` links the same guide.\n\nDone when you have read the reference for the kind and, for an image or a\nvideo, the guide for the model you will use. Each is read once per\nconversation.\n\n### 3. Write the prompt\n\nFollow the structure the reference gives for the mode, in the order the\n`prompt_guide` prefers. Where the two differ, the `prompt_guide` wins: it\nbelongs to the model.\n\nWhen you cannot follow a model's guide (you have not read it, or the user\nhanded you a one-line brief to build on), pass `enhance_prompt: true` on\n`generate_image` or `generate_video` instead: Sceneform rewrites the prompt with that\nmodel's playbook before it generates, keeping every `@{ref:…}` mention. A\nprompt you wrote from the guide needs no rewrite, and a prompt the user asks\nyou to send as is goes unchanged, without `enhance_prompt`.\n\nDone when every line of the reference's checklist holds for this prompt, you\npassed `enhance_prompt: true`, or the user asked for their prompt as is.\n\n### 4. Generate\n\n- Each kind has its own tool: `generate_image`, `generate_video`,\n `generate_speech`, `generate_music`, `generate_sound_effect`,\n `change_voice`, `lip_sync`, `generate_avatar`, `upscale_image`,\n `remove_background`, `reframe_image`. Each takes `requests`, a list of\n what to make.\n- One prompt: one request. Variations of the same image or video prompt are\n one request with `amount` (up to 4).\n- Different prompts or references of one kind: one call with up to 12\n requests. Each other kind is its own call.\n- A set that must match (one product in several scenes, a series): follow\n the reference's set rules, which lock one visual description across every\n prompt.\n- References are refs Sceneform returned: an output's `asset_ref`,\n `search_library` results, a `persona_ref`, `product_ref`, or\n `element_ref`. Where the prompt uses one, write `@{ref:<ref>}` with the\n exact ref you pass (`@{ref:media_abc123}`); Sceneform swaps in the token\n the model reads. A mention of a ref you did not pass is refused. Files attached in the conversation go in `attachments`. A\n pasted link goes through `import_media`. A file on the user's device\n comes in through `open_upload`; in a host without Sceneform cards, follow\n its `without_card` steps.\n- To file results in a Project or folder, pass `project_ref` or\n `folder_ref` (`list_projects`, `list_folders`; `create_project` and\n `create_folder` make new ones when the user names one that does not\n exist).\n\nDone when every requested item has started, or came back as a quote or an\nerror you relay.\n\n### 5. Follow the result\n\nA result with `card_offered: true` was given to the host as a Sceneform\ncard, which shows progress and the finished media. Sceneform cannot see\nwhether the card is on screen. Say it is on its way and stop. When a later\nstep of the request needs the output (an image to animate next), or the user\nasks for the file or cannot see the result, call `wait_operations` with its\n`operation_ref`. Without the flag, call `wait_operations` (it holds up to 25\nseconds per call). Call again only\nfor operations still queued or running, or an entry whose error is\n`retryable`; `succeeded`, `failed`, `canceled`, and any other error are final,\nso report them. Starting the same job again to check on it charges twice.\n\n### 6. Check and deliver\n\nYou often cannot see the output. Before calling it good, run\n`analyze_media` on its `asset_ref` with a question built from the\nreference's checklist: \"Does the label read MORNING, spelled right? Are the\nhands natural?\" Report what you checked, and say plainly when you checked\nnothing.\n\nWhen the check finds one clear defect, name it and offer the fix from the\nreference's fix table. A fix is a new paid generation: run it once the user\nagrees, or when they already asked for hands-off work or the best possible\nresult. Change only the defect, and stop after two fixes per output.\n\nWithout a card, give each finished `url` as a plain link, not an embedded\nimage, so the user keeps it. Reads such as `get_operation` and progress\nchecks never show a card. Call `show_card` only when the user wants to\nreview something now, with only what belongs together: an earlier result\nthey ask to see again, a before and after, the variants of one request, or a\nset they asked to compare. Never use it to recap the session's outputs; each\nreview point gets its own card. When showing a card, say in one sentence\nwhat it shows and give the file links only when the user asks for a file or\ncannot see the card. Add one sentence naming any choice the user did not\nmake (\"4:5 for the Instagram feed\"). Leave prompts, refs, and model names\nout unless asked.\n\n## Spending\n\nGeneration starts at once and charges the credits it returns.\n\n- Pass `quote_only: true` when the user asks what something costs, and on\n every spend call for the rest of the conversation once they ask you to\n check with them before spending.\n- For a quote (`status: \"awaiting_confirmation\"`), give its summary and\n credits in one sentence, and call `confirm_actions` with its `action_ref`\n only after the user says yes. With `card_offered: true` the host was also\n given a card with Generate and Skip buttons; say they can approve there. Confirming the same\n `action_ref` twice never charges twice.\n- A quote with `shortfall` cannot start yet: relay `shortfall.message` and\n stop. Once the account can pay, the same `action_ref` starts it.\n- In a batch, retry only the item that failed with `retryable: true`, with\n its own tool. Resubmitting the batch charges every item again.\n- A failed generation is not resubmitted silently: a retry is a new paid\n request, so ask first.\n\n## Errors\n\nEvery tool error has a `code`, a safe `message`, and `retryable`.\n\n- `INSUFFICIENT_CREDITS`: tell the user and stop. `get_credit_balance` shows\n the balance and `list_credit_transactions` where credits went.\n- `OPTION_UNAVAILABLE` or `PLAN_STALE`: query the options again and call the\n tool again; pass `quote_only: true` again if the user was still deciding.\n- `NOT_FOUND` from `confirm_actions`: the quote expired or belongs to\n another account. Quote again.\n- `TEMPORARY_FAILURE`: relay the message and any `support_ref`. When\n `retryable` is true, retry once after a pause.\n- A call that timed out may have started: find it with `list_operations`\n before anything else.\n"
}SHA-256 of public snapshot: 94eca213f8f24432b7c04e100a168d25ac6f6c4118e25c484cf6dffb6ba11d7b