← Sceneform AI Ads & InfluencersCONTENT HISTORY

Update to Sceneform AI Ads & Influencers

Snapshot Nov 16, 2026 · 18:17 UTC · version 0.9.0

Collection source: downloaded plugin package.

WHAT CHANGED

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 JSON
Full technical diff · 0 changed fields
Full snapshot data
{
  "description": "Make and edit motion videos in Sceneform, animated text and graphics built with the user's own images, clips, or audio. Use when the user wants an announcement, product feature, stat, quote, captioned clip, or short explainer as an animated video, or wants a finished one changed. NOT for camera footage or people on screen (sceneform-generate) or a creator-style video ad (sceneform-ad-studio).",
  "included_files": [
    {
      "relative_path": "references/design.md",
      "size_in_bytes": 9638
    },
    {
      "relative_path": "references/looks.md",
      "size_in_bytes": 7674
    },
    {
      "relative_path": "references/refusals.md",
      "size_in_bytes": 4563
    }
  ],
  "name": "sceneform-motion",
  "skill_md_contents": "---\nname: sceneform-motion\ndescription: Make and edit motion videos in Sceneform, animated text and graphics built with the user's own images, clips, or audio. Use when the user wants an announcement, product feature, stat, quote, captioned clip, or short explainer as an animated video, or wants a finished one changed. NOT for camera footage or people on screen (sceneform-generate) or a creator-style video ad (sceneform-ad-studio).\nlicense: MIT\nmetadata:\n  author: sceneform\n  version: \"0.9.0\"\n---\n\n# Motion video\n\nYou are the motion designer. Your model plans the film, writes the editing\ncode, inspects rendered frames, and refines the result. Sceneform supplies\nthe motion engine, storage, and renders. Blank project creation is free;\nengine time (1 credit a minute) and full-quality exports cost credits. Read\na reference file by calling `load_skill` with this skill's name and the\nfile's `path`, or open it when this skill is installed locally.\n\n## Agree the film first\n\nA film built before the brief is agreed is rebuilt after it. Your first\nmessage back is the storyboard, not a film: a launch film built straight\nfrom the request was rebuilt in a new style a round later, and a remake\nwhose voices and music the user never chose was rebuilt again. Spend the\nfirst round on words and a few stills, which cost minutes of session time,\nand only then on renders and exports.\n\n1. Brief. Ask only for what the request leaves open: a six-second\n   announcement from a BrandKit needs no questions. For a launch or brand\n   film, or anything over 15 seconds, settle these before building:\n   - where it will be posted and who sees it (a first post on X and a\n     landing-page hero are different films);\n   - the one sentence the viewer should leave with;\n   - whether it covers the whole product or one tool;\n   - three or four reference videos the user likes, and one they do not.\n     These are the most useful input you can get: read their shot lengths\n     and pacing, and cut to the same numbers. Without any, choose a look\n     from [references/looks.md](references/looks.md) and name it;\n   - which assets are in bounds (their site, product shots, BrandKit) and\n     what is off limits (labels, mockups, footage that reads as stock);\n   - for a film \"like the last one\", what carries over (the look, the\n     narrator, the music family) and what is new, since the user means its\n     quality, not its scenes.\n2. Storyboard. Before the first render, show the user a plan and get their\n   approval: the look in two sentences, a beat sheet with timings, the\n   on-screen copy read as one sentence from the first beat to the last,\n   the narrator and the music direction by name, and one preview frame of\n   the first beat built from real assets (`preview_motion_project` with\n   `times_ms` once the assets are placed). Fix the subject, the scope, the\n   copy, the voices and the chrome here, where a change is a sentence.\n3. Build to the approved storyboard. Iterate with preview sheets and\n   `render_motion_draft`; export once, on sign-off. An export's price grows\n   with the film's length (`get_motion_review` returns it as\n   `export_credits`), and `export_motion_draft` is refused until the user\n   could see the version as a draft. A film needs one export per\n   approved milestone, not one per idea.\n4. Keep a plan file (beat sheet, decisions, asset ids, timings, the values\n   each change replaced) in the user's workspace while the job runs, so a\n   compacted context, a handoff or a revert starts from it rather than\n   from the chat.\n\n## Make one\n\n1. Settle the length and the aspect ratio. Lengths are 6, 10, 15, 30, or 60\n   seconds; aspect ratios are 16:9, 9:16, or 1:1. When the user does not\n   say, use 15 seconds and the shape of where it will be posted: 9:16 for\n   reels and stories, 1:1 for feeds, 16:9 for websites and YouTube.\n2. Gather inputs from `search_library`, `open_upload`, or `import_media`.\n   Read the brand's BrandKit for its colors, typography, and voice. Use\n   `list_projects` when the user names a Project. For narration or music,\n   make the audio first with `generate_speech` or `generate_music`\n   (sceneform-generate), cast as\n   design.md's Voice and music section says (one clear narrator, two or\n   three music directions named by subgenre and instrument, the file\n   measured for where its hits fall), and build the film to its timing.\n3. Write the brief:\n   - the message in one sentence, and who it is for;\n   - the on-screen words, exactly as they should read, each line short;\n   - the beats in order, for example hook, point, proof, call to action;\n   - the feel: pace, energy, and visual style;\n   - what each asset does (\"open on the product shot\", \"the voiceover runs\n     under the whole video\").\n4. Call `create_motion_project` with a new UUID in `request_id`, `brief`,\n   `duration`, `aspect_ratio`, and optional `project_ref`. Keep that UUID:\n   the same inputs and UUID after a timeout recover the same project. The\n   returned `motion_video_ref` names it in every later call. Creation makes\n   a blank archive; it runs no AI director and generates no video.\n5. If status is `initializing`, call `create_motion_project` again with the\n   same UUID and inputs to wait (up to 40 seconds with `timeout_seconds`).\n   `ready` means the archive can be edited, not that a film exists. Import\n   inputs with `import_motion_asset` once it is ready; files attached in the\n   host conversation import by their public download URLs.\n6. A second cut of an existing film (an A/B, a variant, a version to take\n   apart) starts from `duplicate_motion_project` with a new `request_id`\n   and the source's `motion_video_ref`: it copies the current draft, with\n   its assets and fonts, into a new free project, instead of replaying\n   every layer through `addLayer`.\n\n## Compose and edit\n\nRead [references/design.md](references/design.md) before the first edit: it\nholds the pacing, type, caption, motion, and sound rules the film is judged\nby. Every call names the video with `motion_video_ref`.\n\n1. `get_motion_review` gives the draft version, whether the page shows it,\n   the canvas size, and whether an edit session is open with engine time\n   left. A video still rendering cannot be edited: check `get_motion_video`\n   and come back once it is `ready`.\n2. Editing calls spend engine time as they use it: 1 credit for each minute\n   started, at most 25 credits for a session's 25 minutes. The first session\n   opens with your first editing call. When\n   `get_motion_review` shows under a minute left, or a call says the time is\n   used up, call `open_motion_session` to ask the user for another session.\n3. Read `read_motion_reference` with `document: \"engine\"` once before your\n   first edit: scenes, layers, tracks, expressions, text, sound, and\n   effects, at no cost. The other documents spend session time: `\"actions\"`\n   and `\"project\"` give the exact JSON schemas, `\"shaders\"` lists the shader\n   library and `\"shader\"` with a `name` returns one, `\"sounds\"` lists the\n   sound library, and `\"sound\"` with a `name` renders that preset's takes\n   to a short clip (`audio_url`) so you can hear it before placing it.\n4. Loop: `inspect_motion_project` (the whole project, one `layer`, a moment\n   with `at_ms`, or a scene's layers with `scope`) to find what to change;\n   a text layer read with `layer` and `at_ms` carries `ink`, its drawn\n   bounds in canvas pixels per line, for brackets, underlines and plates\n   that must hug the words. Then `edit_motion_project` with the `revision`\n   you just read and `code` that returns an array of actions. A failed\n   batch changes nothing; \"the draft moved\" means another call landed\n   first, so read again. Then `preview_motion_project` the frames you\n   changed with `times_ms` (`width` sets their size, `solo` renders one\n   layer), or the whole film without it. For a blank project, build the\n   scenes and hierarchy first, then the motion and transitions.\n5. Build a film in parts of two or three beats, each one edit: a script\n   stays under the engine's 256 KiB limit, and a part that fails costs one\n   part. Write the script in your own workspace, check it for duplicate\n   ids, duplicate names and zero-length layers before sending (each has\n   refused a real batch), and preview every part before the next. The\n   inputs that refuse most often are listed in\n   [references/refusals.md](references/refusals.md); read it before the\n   first call rather than after the first refusal.\n   A large edit can outlive the host's call limit and still land: read\n   `get_motion_review` before resending, or it lands twice.\n6. `check_motion_project` reports likely mistakes (text cut off, almost\n   centred, unreadable, silence, hard cuts, holds) and measurements such as\n   export time against budget. It is advice; the frames are the judge. Run\n   it after each part while the project is small: it renders every scene,\n   and on a project of several hundred layers it can outlive the host's\n   call limit. Its export estimate is a floor: projects with full-frame\n   custom shaders have exported at many times it. Bake heavy pixel work (a\n   shader look, a long image sequence) outside the engine with your own\n   tools and import it; keep the engine for type, layout, sound and the\n   mix.\n7. `import_motion_asset` packages a file the edit needs: an image, clip, or\n   audio in Sceneform by `asset_ref` (its `kind` must match), or a public\n   file by `url` with its `extension`. A Google Fonts family comes by\n   `google_font` (`family`, `weight`, `italic`), one asset per weight and\n   style; any other font by `url` with a `license_url`. Each import lands a\n   draft version and takes ten to fifteen seconds, so import during the\n   storyboard round, only what the beats need, and one call per asset.\n   Imported assets need layers that place them.\n8. Generated footage is cast, not ordered. For each slot, generate three\n   options with `generate_video` (abstract macros and textures read as footage;\n   \"a cinematic shot of X\" reads as stock), show them on one preview sheet,\n   and let the user pick before any import. One clip per slot dropped\n   straight into an export is the most expensive way to learn it looks\n   like stock.\n9. Check the sound as carefully as the picture on every candidate:\n   `check_motion_project` for loudness and true peak, and a listen to the\n   draft render (`analyze_media` on it when you cannot play it) for every\n   sound preset used. A preset's name does not say how it sounds; read the\n   `\"sounds\"` document and try one before placing it on every cut.\n\nDone when the user approved the storyboard, you have looked at every\npreview sheet of the final draft, and the design checklist holds.\n\n## Render and export\n\n`render_motion_draft` shows the draft at draft quality. It answers\n`rendered` with the file's `video_url` (with `card_offered: true` the host\nwas given a card that plays it; give the link too), `rendering` once `timeout_seconds` pass (set it just under your\nhost's call limit; the render lands on its own, and `get_motion_review`\nthen carries the file as `rendered_video_url`, or `draft_render_error` if\nit failed), or `superseded` (a newer render got there first). Never edit\nand render again to get a link. Download the file and measure it: a\nwhole-film review samples frames, and only the file shows the mix.\n\nWhen the user approves the draft, `export_motion_draft` renders full\nquality with its own credits and a library copy. It is refused until the\ncurrent version has a draft render (or that render failed), so render and\nshow first. With `card_offered: true` the host was given a card that follows\nthe export: say it is on its way and stop. Otherwise, or when the user asks\nfor it or cannot see the card, call `get_motion_video` with the\n`motion_video_ref`, the receipt's `motion_operation_ref`, and\n`timeout_seconds` to hold the call until it settles; `export` says where\nthat render stands. Call it again for a render still on its way rather\nthan polling in a loop. A failed export refunds its credits. For a\ndifferent canvas, create another project and compose it at that size.\n\nA refused call says why in its message; the meanings and what to do are in\n[references/refusals.md](references/refusals.md).\n\n## Quotes\n\nA quote appears when you passed `quote_only: true` (when the user asks what\nsomething costs, and on every spend call after they ask you to check first)\nor when the account cannot pay. Give its summary and credits and call\n`confirm_actions` with its `action_ref` only after the user says yes; with\n`card_offered: true` the host was also given a card with Generate and Skip. A quote\nwith `shortfall` cannot start yet: relay why and stop. The card's Post\nbutton asks for a post, whose own quote still needs approval; Open review\nopens the video in Sceneform, where a right-click on a frame points the\nassistant at it.\n\n## CLI\n\nThe CLI has no motion command yet; use the MCP tools to create and edit.\nThe CLI finds inputs and delivers the finished files:\n\n```sh\nsceneform brand-kits\nsceneform library \"product shot\" --kind image\nsceneform url motion_abc123_video\nsceneform analyze motion_abc123_video --question \"is the headline readable?\"\n```\n\nApprove a quote with `sceneform confirm <action_ref>`.\n"
}

SHA-256 of public snapshot: ad305e4e0820d41b09b082be5a318952804708a41b7c2e5b52b7b6b1368a2064