ghostfeed-ugc-reactions

Translate the user's goal into the necessary tools. A request to “clone this video” can mean copying its composition, recreating its actions from analysis, or guiding a new performance with the original media. Choose the relevant parts; there is no mandatory end-to-end sequence or MCP App wizard.

Choose the creative operation

Each creative call produces one output. Several references describe one output; they are not a batch count. For several requested outputs, submit distinct calls with distinct idempotency keys and track each generation.

Workspaces, models, and existing assets

Call get_me and list_workspaces when establishing context. Pass the chosen workspace slug or ID explicitly on every write. A Brand name and a workspace name are different identities; resolve each through its own discovery tool.

Use list_image_models and list_reaction_video_modes to choose supported model IDs, reference types/counts, duration, resolution, audio behavior, and prices. Do not invent models or silently drop incompatible references. When a requested combination is unsupported, explain the specific constraint and choose an appropriate alternative with the user.

Reuse a suitable existing image when the user's request permits it. Use list_reaction_frames and get_image to inspect it. Changing the video model or action does not inherently require generating another image. An old frame's existence does not prove that the user liked it; respect their stated choice and any request to review the image before proceeding.

For avatars, use existing avatar discovery and creation tools. Load the ghostfeed-avatars skill when creating a new identity; its approval rules remain unchanged. Do not create a new avatar just to change a scene or garment.

Brands and Product images

• list_brands / get_brand find saved marketing context and processing status.
• create_brand creates a Brand from website information or explicit details. Website enrichment and Product image harvesting are asynchronous; inspect get_brand rather than assuming the images are ready immediately.
• update_brand changes supported details.
• list_product_images returns the Brand's directly owned Product image IDs.
• To add files to a Brand, call request_product_image_upload, PUT bytes to its signed URL with the returned Content-Type, then call add_product_images with the returned upload keys and filenames. Repeat the same ordered inputs when retrying completion. Generic media uploads are a separate reference path.
• Rename or remove a Product image only as requested, using the corresponding Product image tool. Collections are not required for UGC product references.

clone_image selects the product-aware default whenever productImageId is supplied. It inserts that product even when the composition has none, allowing necessary pose/hand/placement changes while preserving avatar identity and product appearance. Template + product cloning without an avatar is unsupported. For product-only scenes, use generate_image with instructions and supported references instead.

Source videos and uploads

Use the source the user supplied or selected. When the request needs a source and none was supplied, discover appropriate sources with list_reaction_templates and list_inspiration_reactions. Library firstFrameScene and the saved videoUnderstanding help selection. Older rows may expose opensOn and motion summaries instead; those are not full action prompts. Import a social link or inspiration into the workspace with import_reaction_template when a saved template is needed. Poll its generation. Long completed imports remain usable. Crop only when the chosen generation workflow needs a shorter or cleaner reference range; render_source_crop supports visual selection in an Apps host.

Choose the part to recreate

After importing a source for cloning, establish which part the user wants before preparing generation. A long organic video often combines a short creator hook, a product demo, and sometimes a creator return. Do not assume the entire clip is the intended motion reference. If the request does not already identify the part, ask one concise question: the opening hook, the creator parts, or the whole video? Respect an explicit selection; do not ask again or force a hook-only workflow.

Read get_reaction_template for source media, duration, and videoUnderstanding. Imports create understanding at no user credit cost and remain processing until it completes. Its saved result contains understanding.firstFrameScene, generationPrompt, description, contentStructure, demoType, simple timestamped sections, and optional visual format, editing complexity, and approximate transcript. New cropped templates are exact media assets with their own understanding and crop-relative timestamps. Historical crops may also carry sourceContext for the whole original and range for the selected source interval. Use the cropped asset's sections to identify the requested part; whole-source categories remain supporting context and never restrict generation. Cropping a template currently starts a new understanding job for that cropped asset. The crop generation can succeed before that analysis finishes: read the new template and wait for its understanding to complete before using its crop-specific descriptions or generation prompt. Do not call analyze_video again for the same crop unless its saved understanding is missing or failed and the user requests a retry. The new analysis describes the crop's actual first frame and action; the whole source's first-frame scene and generation prompt must not be presented as facts about a later segment.

If understanding is absent on an older template, call analyze_video with profile: "understanding"; this explicitly enrolls that template in the same saved analysis lifecycle. For pending or processing, wait briefly and poll the template read or the same analysis request; both reuse the saved job. For failed, explain that analysis failed and use retryFailed: true only when a retry is explicitly requested. Do not loop retries. A new import remains unavailable until its understanding succeeds; older completed templates without this analysis retain their established generation fallback. The 120-second import limit still applies.

Sections and transcripts are approximate, not verified cut points or subtitles. When a boundary matters for motion control, inspect the source or use get_source_frame near the transition and confirm the crop preview. A still frame cannot verify speech or lip-sync. Do not invent timestamps. If the result and visible footage disagree, use the footage and explain the uncertainty.

• Reference-guided motion: isolate the selected source range with crop_reaction_template when the reference includes unwanted demo or other footage. Use the user's supplied/approved range, poll completion, and pass the new template ID as the reference. Keep source audio aligned to the same range. An exact Crop request from the Apps selection authorizes that crop: use its supplied source and start/end times without asking for the same approval again. Use saved understanding and source inspection to choose one continuous range containing the complete requested part. This creates one clip, not an automatic collection of scene clips.
• Full-length reference clone: when the user wants the whole source and a selected model limits reference duration, plan an ordered set of adjacent ranges that covers the requested interval. For Wan 3 Video Guided, each source reference must be at most 15 seconds. Wan 3 and MiniMax H3 Smart Motion use the measured source duration for the output, rounded up to a supported whole second. Wan 3 has a two-second minimum; H3 has a four-second minimum. Do not send durationSeconds for guided or motion-control calls; supplied values are rejected before billing. For a faithful clone, plan each range near a supported whole-second length. If a natural boundary leaves a fractional duration, adjust a nearby cut when the footage allows it, or explain the duration mismatch before generation. Place boundaries at natural changes in movement or scene, not mechanically at 15-second marks. Use the original understanding to propose boundaries, then inspect nearby source frames and the crop preview; its timestamps are approximate. Avoid an impractically short final range by adjusting an earlier boundary. Crop each range from the original source, poll each crop generation, and use its own starting frame/composition for the selected avatar. Generate one video per crop in source order, with separate idempotency keys. Do not silently switch to another model or offer one 15-second reference as a faithful clone of a longer performance. Price the complete plan, including any additional starting images, before paid generation. The outputs are independent takes: inspect transitions and do not promise seamless motion or audio. MCP currently has no video-stitching tool; provide the dashboard editor link for assembly and say clearly when the requested final combined video has not yet been delivered.
• Prompt-directed recreation: read the source understanding, then write the prompt for only the selected section yourself. The backend does not write or insert the source's saved generation prompt for this tool. Cropping the original is unnecessary when only written action guides generation. If there is not enough evidence to write that action, resolve the missing detail before generation.
• Demo footage: ask for or reuse the user's own product demonstration when it is needed in the result. Do not copy the source app's interface or present invented functionality as the user's product. A hook-only request needs no demo. For assembly beyond generation, provide the dashboard editing link.

These choices also apply to animation and other source formats. A category helps interpret the footage; it does not decide what can be generated or override the user's goal.

For local Image, Video, or Audio references:

1. request_media_upload with filename, Content-Type, byte length, and a stable upload ID when retrying.
2. PUT the bytes to the returned signed URL using its Content-Type.
3. complete_media_upload with uploadId. Use the returned mediaId thereafter.

An upload URL alone is not a usable reference. Completion validates media and stores an immutable accepted asset with server-measured audio/video duration.

Sources for analysis use exactly one of { "templateId": "..." }, { "mediaId": "..." }, or { "videoId": "..." } for a completed generated video in the selected workspace. Reuse that video ID directly; downloading and reuploading the generated video is unnecessary. Frame extraction accepts only { "templateId": "..." } or { "mediaId": "..." }. get_source_frame accepts source and timestampSeconds and returns a saved frameId. It works directly on completed uploaded videos; template import is not required. Repeating the source/timestamp reuses the saved frame. Audio-only uploads cannot supply an image frame or visual action analysis.

Image references use explicit kinds, for example { "kind": "frame", "frameId": "..." } or { "kind": "upload", "mediaId": "..." }. Use returned IDs rather than inventing URLs or translating one kind of ID into another.

Images and composition

generate_image supports text-only creation and multiple references without an avatar when the selected model supports them. For identity or product-led creation, supply the corresponding IDs and a concise scene instruction.

clone_image accepts a template, saved frame, or completed uploaded image as composition. It requires an avatar and automatically chooses the applicable clone default. To copy a specific moment in a video, first obtain its saved frame with get_source_frame, then use that frame as the composition.

edit_image makes a new image from a saved frame or completed uploaded image and a requested change. The edit default preserves unrequested content. Follow the user's requested review checkpoints; the tools do not require presenting every system prompt before execution.

Two ways to recreate a video

Action described by a prompt

Use generate_video_from_prompt when written action controls the result. Supply the complete prompt yourself. For a source recreation, read its saved videoUnderstanding and inspect the requested section, then write a prompt for that result. The saved generationPrompt is context, not text the backend silently submits. For older sources without understanding, use analyze_video to obtain source context. If the user supplies an exact prompt, preserve it.

The source supplies action text; this does not automatically make its video a motion-control input. The starting image's composition source and the action source can differ. Your prompt must not inherit unrelated source movement.

Performance guided by reference media

Use generate_video_from_reference when motion Video or Audio directly guides the output. A completed generated video can be the reference as { "videoId": "..." }; in the additional references array use { "kind": "video", "videoId": "..." }. Supply the starting image and explicit reference; it does not need to be the image's original source. Existing motion-control modes are one_to_one_standard and one_to_one_clone_premium. Their supported input and duration rules come from the catalogue; do not promise identical generated pixels or unrestricted duration control.

For direct video cloning, recommend Wan 3 Smart Motion — Video Guided by default: generate_video_from_reference with mode: "wan_3" and guidance: "video_guided". Check its reference limits and pricing first. Respect an explicitly selected model; when Wan 3 cannot support the requested input, explain the constraint and choose a supported crop or alternative with the user. For audio-led performance, use Wan 3 Smart Motion — Audio Guided when appropriate.

Use dashboard model names in conversation: Kling 2.6 Motion Control maps to one_to_one_standard; Kling 3 Motion Control maps to one_to_one_clone_premium. These are compatibility IDs, not user-facing tier names. Do not describe these models as “1:1 Standard” or “1:1 Clone Premium”.

For “clone this video,” select the family that fits the user's requested result. Prompt recreation suits adapting the action. Reference guidance suits directly following source motion/performance. Ask about this difference only when it materially changes the intended result and the request does not resolve it.

For a visible lip-sync performance to an existing sound, prefer supported video-guided motion when the user wants to keep that performance. Use a clean reference range and preserve the matching original audio when the selected mode supports it; check the catalogue's actual audio behavior. A transcript is context, not a replacement for the sound's delivery and timing. Music under silent reactions or B-roll does not establish lip-sync. If overlays obscure the performance, obtain a usable reference or explain the limitation. Do not promise exact mouth synchronization. New words or a new voice may need a different path.

Prompt-directed H3 and reference-guided Wan 3

Do not recommend H3 Smart Motion for new work. Its guided modes remain accepted temporarily for compatibility with existing clients and saved setups. Use H3 or H3 Max for prompt-directed action and Wan 3 for new audio/video-guided work.

Both Smart Motion choices use one starting image plus one guidance source. Do not attach extra product images to guided modes; incorporate the product into the starting image first if needed. Ordinary H3 has its own additional-reference capabilities, and H3 Max has its own model settings.

Audio Guided: the reference may be Audio or a Video whose audio the backend extracts. Video sources receive the appropriate audio-guided action preparation. Audio-only sources do not require visual analysis or a custom action prompt. The backend selects supported output duration from the source. Duration-dependent action preparation is refreshed only for legacy templates or direct media that lack a saved canonical generation prompt.

Video Guided: supply a Video reference. Wan 3 Video Guided uses the inspectable wan3-video-guided system prompt by default, but needs no custom action prompt: the starting image and source video carry identity and motion. Do not analyze or write an action prompt unless the user asks to change the source performance. For H3 Video Guided, new template imports reuse their canonical provider-authored generation prompt; older templates and direct uploads fall back to motion-prompt analysis. Reference input duration can affect cost. For either guided choice, actionPrompt can explicitly specify the action and adaptation expresses requested changes.

Read get_prompt_template with wan3-video-guided to inspect or edit Wan 3's default system guidance before generation when needed. Passing systemPrompt replaces that default exactly; null removes it. The submitted system text is saved in get_generation_prompt. Other public guided modes do not receive a default system prompt.

Analysis and prompt transparency

analyze_video has two purposes. understanding reads or explicitly starts the free, asynchronous saved section analysis described above. For templates it uses the same canonical record returned by get_reaction_template. Generation action preparation is optional: choose standard, audio_guided, or video_guided; audio-guided analysis requires outputDurationSeconds. The generation tools use the same preparation automatically when source-derived action is requested.

Keep these distinct:

• An image's saved generation prompt.
• A source video's factual understanding, section map, and provider-authored generation prompt.
• A legacy source video's fallback motion-prompt analysis.
• Any explicit caller-supplied system guidance.
• The final submitted prompt and reference order.

A frame may link to a source without storing an independent video action prompt. Use get_image for lineage. Do not substitute today's analysis or defaults when claiming to reproduce a previous generation's actual settings.

The MCP adapter never invents a system prompt. Wan 3 Video Guided selects its inspectable backend default unless the caller overrides or removes it. Submitted system text is saved and can be inspected using get_generation_prompt. Retrieve saved nonbilling settings with get_generation_setup. Historical unavailable fields are not permission to invent missing settings.

Jobs, retries, and delivery

• Agree on the proposed paid work and its estimated cost or a scoped budget before submitting it. Permission to import, inspect, or test is not permission to spend on additional analysis or generation. Respect an already approved scope and ceiling; ask again before expanding them. A retry is not permission to pay for a new output.
• Supply a stable idempotencyKey before each paid request. Retry the identical request with that same key after a transport failure. A changed request or a deliberately new output needs a new key. If preparation ends with preparation_expired, the same key preserves that failure. Create a new key only for an explicitly requested retry, rather than repeatedly resubmitting.
• Poll the returned generation ID with get_generation. Use the returned status and error information; do not resubmit merely because preparation or rendering is slow. Generation success is succeeded, not the board's complete label.
• Use model pricing and maxCredits when a budget ceiling matters. Report actual credits from returned data, not a guessed amount.
• Show images with render_image_results, videos with render_video_result, and history with render_generation_gallery in MCP Apps hosts. Use the supplied IDs, URLs, and generation-specific dashboard links. Apps support inspection and choices; the agent decides the creative workflow.
• Apps buttons may send contextual, open-ended chat requests containing the selected asset and its saved context. Treat a Reference or Recreate selection as the starting point for the conversation: inspect the supplied context and ask only for creative intent that is still missing. Reuse existing answers and explicit instructions; do not make the user restate them. Selecting an asset does not itself approve paid generation. Follow the user's existing spending authorization and review preferences before a paid tool call.
• Without App rendering, use the returned media and dashboard links. Do not claim an asset was generated or visually checked before it is available.

Legacy reaction generation calls remain available during migration. Their notices name the replacement process and the October 15, 2026 retirement date. Use the new intent tools for new work; never repeat a successful paid legacy generation just to migrate its call. Slideshow workflows, finished-video editing, and publishing are outside this skill's UGC generation scope.