Update to Widemile
Snapshot Oct 29, 2026 · 00:05 UTC · version 1.2.0
Collection source: downloaded plugin package.
Payment or plan references changed
Skill instructions — Now says: each requested video sequence as one generated clip when the selected model supports its full duration, references and intended cuts. A sequence may contain...
Observed in published text. Live prices and checkout terms have not been verified by this change.
Up to 3 plugins free · no card required.
All fields and original values.
Skill instructions
Oct 24, 2026 · 13:07 UTC
one video beat as one generated clip. Before preparing it, call `prostudio_models_list` with `mediaType=video` and choose `modelId`; choose `durationSeconds` from what that beat needs except for source-locked Seedance editing, where the ...
Oct 29, 2026 · 00:05 UTC
each requested video sequence as one generated clip when the selected model supports its full duration, references and intended cuts. A sequence may contain multiple beats, shots, cuts or scenes; do not split it merely at a scene change....
Compare saved observations
Download comparison JSONSKILL.md text changes
− removed · + added
--- before +++ after @@ -15,10 +15,11 @@ - The Brief is optional creative context, never a technical prerequisite. An empty Brief is valid: do not create, save, expand, or request approval for a Brief merely to unlock `prostudio_generation_prepare`. Prepare directly from the user's request, live model capabilities, and explicit asset bindings. - Apply the host's model-selection policy before planning a generation. Before any new model-selectable media generation, call `prostudio_models_list` for the target media type and select from the live catalog. An explicit user model choice wins. Do not silently apply a default model when the user asked for a recommendation, maximum quality, lowest cost, or another exact available model. -- Outside storyboard-driven production, treat one video beat as one generated clip. Before preparing it, call `prostudio_models_list` with `mediaType=video` and choose `modelId`; choose `durationSeconds` from what that beat needs except for source-locked Seedance editing, where the host derives it from the supplied video. When the model changes, rebuild generation mode, resolution, duration, aspect ratio, audio behavior, and reference bindings from the new capability; never carry model-dependent parameters forward. +- Outside storyboard-driven production, treat each requested video sequence as one generated clip when the selected model supports its full duration, references and intended cuts. A sequence may contain multiple beats, shots, cuts or scenes; do not split it merely at a scene change. Before preparing it, call `prostudio_models_list` with `mediaType=video` and choose `modelId`; choose `durationSeconds` from what that beat needs except for source-locked Seedance editing, where the host derives it from the supplied video. When the model changes, rebuild generation mode, resolution, duration, aspect ratio, audio behavior, and reference bindings from the new capability; never carry model-dependent parameters forward. - For Seedance 2.5 with video inputs, choose `options.seedanceReferenceOperation`: `generation` uses references to create a new video; `editing` modifies one source; `extension` continues one source. Resolve the references first and set `options.seedanceSourceVideoIndex` for editing/extension to the 1-based source in the ordered video bindings. Name that exact `@videoN` in the prompt and explain how other inputs support it. Never assume the first input is the source or concatenate input videos implicitly. At most 10 videos are supported, within 30 seconds combined. Editing requires every video input to be 4–30 seconds, preserves the selected source's ratio and actual duration (including fractional seconds), and requires omitting `durationSeconds` and `aspectRatio`. If the user requests another duration or ratio, plan a separate trim or reframing operation. Extension preserves the source ratio; its selected 4–30 second duration is the new output clip's duration, not the source duration plus an increment. Reference generation supports independent output duration and ratio. - Seedance 2.5 supports `options.draft: true` for a 480p preview. Draft creation and final rendering always use KIE. Use Draft Mode only for the generation operation (text, image, or reference generation); source-locked editing and extension require standard mode. Quote the draft separately; do not claim that the initial credits include the final render. The user can render the same draft at 1080p within 7 days using the final-render icon on its Widemile card and confirming the displayed credits. Do not regenerate from the draft video as a reference to simulate final rendering. -- Generate audio for video by default. For always-on-audio models, omit `options.generateAudio`. For configurable audio, set it to true unless the user explicitly requests silence; then select a compatible configurable model and disable it. Never describe an always-on-audio model as silent or audio-incapable. +- For Kling 4, use only `kling-4.0/reference-to-video`, with optional image/video references and tokens `@image_1`/`@video_1`. Output: 3–30 integer seconds, 720p/1080p, auto/16:9/9:16/1:1/21:9. Files: 10 MB each; at most 5 videos, 2–30 seconds each and 30 seconds combined. Use `options.keepOriginalSound` (default false) for source video sound; omit `generateAudio`. Elements, voices, uploaded audio and frame mode are unavailable. +- Generate audio for video by default, except Kling 4 original-sound retention as described above. For always-on-audio models, omit `options.generateAudio`. For configurable audio, set it to true unless the user explicitly requests silence; then select a compatible configurable model and disable it. Never describe an always-on-audio model as silent or audio-incapable. - Every video plan outside source-locked Seedance editing must declare an integer `durationSeconds` supported exactly by its selected product model. A final deliverable duration is an editing goal, not one generation: split longer work into independent clips and assemble them in Editor. `shotId`, when supplied, is only a concise optional workflow label and does not need to exist in the Brief. - Read the human Brief for creative intent, then select exact stable references from the live reference library and bind their purpose explicitly in the generation plan. Never parse reference ids, model ids, duration, or workflow state from Markdown, and never substitute unrelated or placeholder media. Keep generation prompts in English while preserving quoted dialogue, voice-over copy, lyrics, visible text, proper names, and other exact content in the requested target language. @@ -31,3 +32,5 @@ - Bind the selected shots' completed image `frameGenerationId` values through `continuityGenerationIds`, deduplicated in first-cut order. These are generation IDs, not reference-library IDs. Copy scene, shot and generation UUIDs exactly from the latest storyboard response; never shorten or reconstruct them. Before preparation, verify every cut-to-source mapping against the resolver output. If resolution fails, re-read the canonical storyboard and compare and recopy its exact source IDs before retrying, rather than repeating an unverified ID or changing the image. Resolve the full scene reference list successfully before using its tokens. Never merge token maps from individually or partially resolved subsets: token positions may change in the full list. If it remains unresolved, report the specific scene as blocked; do not present guessed bindings as a ready prompt. Other scenes can continue. Resolve them with `prostudio_references_resolve`, retain every selected image, and associate each cut with the actual binding token returned for that source ID. A reused image can guide several cuts with the same returned token. Add other references only when they serve the requested scene. - A cut without a selected image remains valid textual direction; do not force image or reference creation. Choose a model that supports the full scene's duration, selected references and intended cut sequence. Preserve each specified shot duration and give every cut a positive time interval in the prompt. The cut intervals must be ordered, contiguous, non-overlapping and cover the full scene. When every cut is timed, their sum is the scene duration. When some cuts are untimed, reserve positive time for each of them: the scene duration must exceed the sum of the specified durations, and added timings must be identified as estimates. Never assign zero remaining seconds to an untimed cut or shorten a specified cut silently. Choose a supported total that preserves the scene rhythm; if no exact compatible duration exists, report the limit before preparation. If the chosen model cannot accommodate the scene, explain the specific limit and agree on a compatible model or scene revision before paid preparation; do not silently omit images or split the scene into separate per-shot videos. Unrelated production work can continue. - Follow the normal prepare, exact-price consent and execution flow for each requested scene. When recording its completed video, use `prostudio_shot_generation_set` with `shotId` equal to the scene UUID and `generationId` equal to the completed video. This association stays in operational state; the storyboard shots' `frameGenerationId` fields continue to point to their chosen image references. Editing a storyboard alone never requests video generation. + +- Billing display is controlled by the authenticated account. When `billingDisplay.unit` is `EUR`, present the returned euro charge and balances; otherwise present credits. Preserve prepared execution arguments and `expectedCredits` exactly, without displaying those machine values as prices.
Full technical diff · 1 changed fields
changed /skill_md_contents
"---\nname: generation-planning\ndescription: Select available product models and prepare independent image, video, audio, or music plans with exact duration, audio, and reference constraints.\n---\n\n# Generation planning\n\n## Tool requirements\n\n- **Core planning:** `prostudio_models_list` and `prostudio_generation_prepare`.\n- **Conditional creative context and bindings:** `prostudio_brief_get`, `prostudio_workspace_get`, `prostudio_references_list`, and `prostudio_references_resolve` when the generation depends on the Brief or references.\n- **Conditional storyboard-driven video:** `prostudio_storyboard_get` and `prostudio_references_resolve`; use `prostudio_shot_generation_set` when recording a completed scene video.\n- **Conditional execution and verification:** exactly the required dispatch tool among `prostudio_image_generate`, `prostudio_video_generate`, `prostudio_audio_generate`, and `prostudio_music_generate`, followed by `prostudio_job_status` and `prostudio_generation_get`. Use `prostudio_generation_plans_list`, `prostudio_generation_plan_get`, or `prostudio_generation_plan_cancel` only for existing plan management.\n- This declaration does not activate tools. Use only the tools required by the selected branch. If a required tool is absent from the current scoped catalog and `load_prostudio_tools` is available, call it with the missing outcome before claiming the capability is unavailable. If retrieval still exposes no matching tool, report that limitation. A model-catalog lookup alone is never completion for a requested generation.\n\n- The Brief is optional creative context, never a technical prerequisite. An empty Brief is valid: do not create, save, expand, or request approval for a Brief merely to unlock `prostudio_generation_prepare`. Prepare directly from the user's request, live model capabilities, and explicit asset bindings.\n- Apply the host's model-selection policy before planning a generation. Before any new model-selectable media generation, call `prostudio_models_list` for the target media type and select from the live catalog. An explicit user model choice wins. Do not silently apply a default model when the user asked for a recommendation, maximum quality, lowest cost, or another exact available model.\n- Outside storyboard-driven production, treat one video beat as one generated clip. Before preparing it, call `prostudio_models_list` with `mediaType=video` and choose `modelId`; choose `durationSeconds` from what that beat needs except for source-locked Seedance editing, where the host derives it from the supplied video. When the model changes, rebuild generation mode, resolution, duration, aspect ratio, audio behavior, and reference bindings from the new capability; never carry model-dependent parameters forward.\n- For Seedance 2.5 with video inputs, choose `options.seedanceReferenceOperation`: `generation` uses references to create a new video; `editing` modifies one source; `extension` continues one source. Resolve the references first and set `options.seedanceSourceVideoIndex` for editing/extension to the 1-based source in the ordered video bindings. Name that exact `@videoN` in the prompt and explain how other inputs support it. Never assume the first input is the source or concatenate input videos implicitly. At most 10 videos are supported, within 30 seconds combined. Editing requires every video input to be 4–30 seconds, preserves the selected source's ratio and actual duration (including fractional seconds), and requires omitting `durationSeconds` and `aspectRatio`. If the user requests another duration or ratio, plan a separate trim or reframing operation. Extension preserves the source ratio; its selected 4–30 second duration is the new output clip's duration, not the source duration plus an increment. Reference generation supports independent output duration and ratio.\n- Seedance 2.5 supports `options.draft: true` for a 480p preview. Draft creation and final rendering always use KIE. Use Draft Mode only for the generation operation (text, image, or reference generation); source-locked editing and extension require standard mode. Quote the draft separately; do not claim that the initial credits include the final render. The user can render the same draft at 1080p within 7 days using the final-render icon on its Widemile card and confirming the displayed credits. Do not regenerate from the draft video as a reference to simulate final rendering.\n- Generate audio for video by default. For always-on-audio models, omit `options.generateAudio`. For configurable audio, set it to true unless the user explicitly requests silence; then select a compatible configurable model and disable it. Never describe an always-on-audio model as silent or audio-incapable.\n- Every video plan outside source-locked Seedance editing must declare an integer `durationSeconds` supported exactly by its selected product model. A final deliverable duration is an editing goal, not one generation: split longer work into independent clips and assemble them in Editor. `shotId`, when supplied, is only a concise optional workflow label and does not need to exist in the Brief.\n- Read the human Brief for creative intent, then select exact stable references from the live reference library and bind their purpose explicitly in the generation plan. Never parse reference ids, model ids, duration, or workflow state from Markdown, and never substitute unrelated or placeholder media. Keep generation prompts in English while preserving quoted dialogue, voice-over copy, lyrics, visible text, proper names, and other exact content in the requested target language.\n\n## Storyboard-driven video\n\nUse this branch when the user asks to develop video from a storyboard or the storyboard is useful to the current production. Read `prostudio_storyboard_get` for the latest relevant scenes. The storyboard remains optional for all other work.\n\n- **One scene = one video generation.** Prepare one independent plan and one complete video prompt for each requested scene. Its ordered inquadrature/shots are the cuts within that same clip. Use the stable scene UUID as the plan's optional `shotId` workflow label.\n- Write the prompt as an explicit ordered cut sequence. For each cut, carry over its action/composition caption, chosen visual reference, camera movement, focal length and timing when specified. Describe the transition into the next cut and preserve subject, setting and visual continuity. Preserve the user's choices and requested style.\n- Bind the selected shots' completed image `frameGenerationId` values through `continuityGenerationIds`, deduplicated in first-cut order. These are generation IDs, not reference-library IDs. Copy scene, shot and generation UUIDs exactly from the latest storyboard response; never shorten or reconstruct them. Before preparation, verify every cut-to-source mapping against the resolver output. If resolution fails, re-read the canonical storyboard and compare and recopy its exact source IDs before retrying, rather than repeating an unverified ID or changing the image. Resolve the full scene reference list successfully before using its tokens. Never merge token maps from individually or partially resolved subsets: token positions may change in the full list. If it remains unresolved, report the specific scene as blocked; do not present guessed bindings as a ready prompt. Other scenes can continue. Resolve them with `prostudio_references_resolve`, retain every selected image, and associate each cut with the actual binding token returned for that source ID. A reused image can guide several cuts with the same returned token. Add other references only when they serve the requested scene.\n- A cut without a selected image remains valid textual direction; do not force image or reference creation. Choose a model that supports the full scene's duration, selected references and intended cut sequence. Preserve each specified shot duration and give every cut a positive time interval in the prompt. The cut intervals must be ordered, contiguous, non-overlapping and cover the full scene. When every cut is timed, their sum is the scene duration. When some cuts are untimed, reserve positive time for each of them: the scene duration must exceed the sum of the specified durations, and added timings must be identified as estimates. Never assign zero remaining seconds to an untimed cut or shorten a specified cut silently. Choose a supported total that preserves the scene rhythm; if no exact compatible duration exists, report the limit before preparation. If the chosen model cannot accommodate the scene, explain the specific limit and agree on a compatible model or scene revision before paid preparation; do not silently omit images or split the scene into separate per-shot videos. Unrelated production work can continue.\n- Follow the normal prepare, exact-price consent and execution flow for each requested scene. When recording its completed video, use `prostudio_shot_generation_set` with `shotId` equal to the scene UUID and `generationId` equal to the completed video. This association stays in operational state; the storyboard shots' `frameGenerationId` fields continue to point to their chosen image references. Editing a storyboard alone never requests video generation.\n"
"---\nname: generation-planning\ndescription: Select available product models and prepare independent image, video, audio, or music plans with exact duration, audio, and reference constraints.\n---\n\n# Generation planning\n\n## Tool requirements\n\n- **Core planning:** `prostudio_models_list` and `prostudio_generation_prepare`.\n- **Conditional creative context and bindings:** `prostudio_brief_get`, `prostudio_workspace_get`, `prostudio_references_list`, and `prostudio_references_resolve` when the generation depends on the Brief or references.\n- **Conditional storyboard-driven video:** `prostudio_storyboard_get` and `prostudio_references_resolve`; use `prostudio_shot_generation_set` when recording a completed scene video.\n- **Conditional execution and verification:** exactly the required dispatch tool among `prostudio_image_generate`, `prostudio_video_generate`, `prostudio_audio_generate`, and `prostudio_music_generate`, followed by `prostudio_job_status` and `prostudio_generation_get`. Use `prostudio_generation_plans_list`, `prostudio_generation_plan_get`, or `prostudio_generation_plan_cancel` only for existing plan management.\n- This declaration does not activate tools. Use only the tools required by the selected branch. If a required tool is absent from the current scoped catalog and `load_prostudio_tools` is available, call it with the missing outcome before claiming the capability is unavailable. If retrieval still exposes no matching tool, report that limitation. A model-catalog lookup alone is never completion for a requested generation.\n\n- The Brief is optional creative context, never a technical prerequisite. An empty Brief is valid: do not create, save, expand, or request approval for a Brief merely to unlock `prostudio_generation_prepare`. Prepare directly from the user's request, live model capabilities, and explicit asset bindings.\n- Apply the host's model-selection policy before planning a generation. Before any new model-selectable media generation, call `prostudio_models_list` for the target media type and select from the live catalog. An explicit user model choice wins. Do not silently apply a default model when the user asked for a recommendation, maximum quality, lowest cost, or another exact available model.\n- Outside storyboard-driven production, treat each requested video sequence as one generated clip when the selected model supports its full duration, references and intended cuts. A sequence may contain multiple beats, shots, cuts or scenes; do not split it merely at a scene change. Before preparing it, call `prostudio_models_list` with `mediaType=video` and choose `modelId`; choose `durationSeconds` from what that beat needs except for source-locked Seedance editing, where the host derives it from the supplied video. When the model changes, rebuild generation mode, resolution, duration, aspect ratio, audio behavior, and reference bindings from the new capability; never carry model-dependent parameters forward.\n- For Seedance 2.5 with video inputs, choose `options.seedanceReferenceOperation`: `generation` uses references to create a new video; `editing` modifies one source; `extension` continues one source. Resolve the references first and set `options.seedanceSourceVideoIndex` for editing/extension to the 1-based source in the ordered video bindings. Name that exact `@videoN` in the prompt and explain how other inputs support it. Never assume the first input is the source or concatenate input videos implicitly. At most 10 videos are supported, within 30 seconds combined. Editing requires every video input to be 4–30 seconds, preserves the selected source's ratio and actual duration (including fractional seconds), and requires omitting `durationSeconds` and `aspectRatio`. If the user requests another duration or ratio, plan a separate trim or reframing operation. Extension preserves the source ratio; its selected 4–30 second duration is the new output clip's duration, not the source duration plus an increment. Reference generation supports independent output duration and ratio.\n- Seedance 2.5 supports `options.draft: true` for a 480p preview. Draft creation and final rendering always use KIE. Use Draft Mode only for the generation operation (text, image, or reference generation); source-locked editing and extension require standard mode. Quote the draft separately; do not claim that the initial credits include the final render. The user can render the same draft at 1080p within 7 days using the final-render icon on its Widemile card and confirming the displayed credits. Do not regenerate from the draft video as a reference to simulate final rendering.\n- For Kling 4, use only `kling-4.0/reference-to-video`, with optional image/video references and tokens `@image_1`/`@video_1`. Output: 3–30 integer seconds, 720p/1080p, auto/16:9/9:16/1:1/21:9. Files: 10 MB each; at most 5 videos, 2–30 seconds each and 30 seconds combined. Use `options.keepOriginalSound` (default false) for source video sound; omit `generateAudio`. Elements, voices, uploaded audio and frame mode are unavailable.\n- Generate audio for video by default, except Kling 4 original-sound retention as described above. For always-on-audio models, omit `options.generateAudio`. For configurable audio, set it to true unless the user explicitly requests silence; then select a compatible configurable model and disable it. Never describe an always-on-audio model as silent or audio-incapable.\n- Every video plan outside source-locked Seedance editing must declare an integer `durationSeconds` supported exactly by its selected product model. A final deliverable duration is an editing goal, not one generation: split longer work into independent clips and assemble them in Editor. `shotId`, when supplied, is only a concise optional workflow label and does not need to exist in the Brief.\n- Read the human Brief for creative intent, then select exact stable references from the live reference library and bind their purpose explicitly in the generation plan. Never parse reference ids, model ids, duration, or workflow state from Markdown, and never substitute unrelated or placeholder media. Keep generation prompts in English while preserving quoted dialogue, voice-over copy, lyrics, visible text, proper names, and other exact content in the requested target language.\n\n## Storyboard-driven video\n\nUse this branch when the user asks to develop video from a storyboard or the storyboard is useful to the current production. Read `prostudio_storyboard_get` for the latest relevant scenes. The storyboard remains optional for all other work.\n\n- **One scene = one video generation.** Prepare one independent plan and one complete video prompt for each requested scene. Its ordered inquadrature/shots are the cuts within that same clip. Use the stable scene UUID as the plan's optional `shotId` workflow label.\n- Write the prompt as an explicit ordered cut sequence. For each cut, carry over its action/composition caption, chosen visual reference, camera movement, focal length and timing when specified. Describe the transition into the next cut and preserve subject, setting and visual continuity. Preserve the user's choices and requested style.\n- Bind the selected shots' completed image `frameGenerationId` values through `continuityGenerationIds`, deduplicated in first-cut order. These are generation IDs, not reference-library IDs. Copy scene, shot and generation UUIDs exactly from the latest storyboard response; never shorten or reconstruct them. Before preparation, verify every cut-to-source mapping against the resolver output. If resolution fails, re-read the canonical storyboard and compare and recopy its exact source IDs before retrying, rather than repeating an unverified ID or changing the image. Resolve the full scene reference list successfully before using its tokens. Never merge token maps from individually or partially resolved subsets: token positions may change in the full list. If it remains unresolved, report the specific scene as blocked; do not present guessed bindings as a ready prompt. Other scenes can continue. Resolve them with `prostudio_references_resolve`, retain every selected image, and associate each cut with the actual binding token returned for that source ID. A reused image can guide several cuts with the same returned token. Add other references only when they serve the requested scene.\n- A cut without a selected image remains valid textual direction; do not force image or reference creation. Choose a model that supports the full scene's duration, selected references and intended cut sequence. Preserve each specified shot duration and give every cut a positive time interval in the prompt. The cut intervals must be ordered, contiguous, non-overlapping and cover the full scene. When every cut is timed, their sum is the scene duration. When some cuts are untimed, reserve positive time for each of them: the scene duration must exceed the sum of the specified durations, and added timings must be identified as estimates. Never assign zero remaining seconds to an untimed cut or shorten a specified cut silently. Choose a supported total that preserves the scene rhythm; if no exact compatible duration exists, report the limit before preparation. If the chosen model cannot accommodate the scene, explain the specific limit and agree on a compatible model or scene revision before paid preparation; do not silently omit images or split the scene into separate per-shot videos. Unrelated production work can continue.\n- Follow the normal prepare, exact-price consent and execution flow for each requested scene. When recording its completed video, use `prostudio_shot_generation_set` with `shotId` equal to the scene UUID and `generationId` equal to the completed video. This association stays in operational state; the storyboard shots' `frameGenerationId` fields continue to point to their chosen image references. Editing a storyboard alone never requests video generation.\n\n- Billing display is controlled by the authenticated account. When `billingDisplay.unit` is `EUR`, present the returned euro charge and balances; otherwise present credits. Preserve prepared execution arguments and `expectedCredits` exactly, without displaying those machine values as prices.\n"
Full snapshot data
{
"description": "Select available product models and prepare independent image, video, audio, or music plans with exact duration, audio, and reference constraints.",
"included_files": [],
"name": "generation-planning",
"skill_md_contents": "---\nname: generation-planning\ndescription: Select available product models and prepare independent image, video, audio, or music plans with exact duration, audio, and reference constraints.\n---\n\n# Generation planning\n\n## Tool requirements\n\n- **Core planning:** `prostudio_models_list` and `prostudio_generation_prepare`.\n- **Conditional creative context and bindings:** `prostudio_brief_get`, `prostudio_workspace_get`, `prostudio_references_list`, and `prostudio_references_resolve` when the generation depends on the Brief or references.\n- **Conditional storyboard-driven video:** `prostudio_storyboard_get` and `prostudio_references_resolve`; use `prostudio_shot_generation_set` when recording a completed scene video.\n- **Conditional execution and verification:** exactly the required dispatch tool among `prostudio_image_generate`, `prostudio_video_generate`, `prostudio_audio_generate`, and `prostudio_music_generate`, followed by `prostudio_job_status` and `prostudio_generation_get`. Use `prostudio_generation_plans_list`, `prostudio_generation_plan_get`, or `prostudio_generation_plan_cancel` only for existing plan management.\n- This declaration does not activate tools. Use only the tools required by the selected branch. If a required tool is absent from the current scoped catalog and `load_prostudio_tools` is available, call it with the missing outcome before claiming the capability is unavailable. If retrieval still exposes no matching tool, report that limitation. A model-catalog lookup alone is never completion for a requested generation.\n\n- The Brief is optional creative context, never a technical prerequisite. An empty Brief is valid: do not create, save, expand, or request approval for a Brief merely to unlock `prostudio_generation_prepare`. Prepare directly from the user's request, live model capabilities, and explicit asset bindings.\n- Apply the host's model-selection policy before planning a generation. Before any new model-selectable media generation, call `prostudio_models_list` for the target media type and select from the live catalog. An explicit user model choice wins. Do not silently apply a default model when the user asked for a recommendation, maximum quality, lowest cost, or another exact available model.\n- Outside storyboard-driven production, treat each requested video sequence as one generated clip when the selected model supports its full duration, references and intended cuts. A sequence may contain multiple beats, shots, cuts or scenes; do not split it merely at a scene change. Before preparing it, call `prostudio_models_list` with `mediaType=video` and choose `modelId`; choose `durationSeconds` from what that beat needs except for source-locked Seedance editing, where the host derives it from the supplied video. When the model changes, rebuild generation mode, resolution, duration, aspect ratio, audio behavior, and reference bindings from the new capability; never carry model-dependent parameters forward.\n- For Seedance 2.5 with video inputs, choose `options.seedanceReferenceOperation`: `generation` uses references to create a new video; `editing` modifies one source; `extension` continues one source. Resolve the references first and set `options.seedanceSourceVideoIndex` for editing/extension to the 1-based source in the ordered video bindings. Name that exact `@videoN` in the prompt and explain how other inputs support it. Never assume the first input is the source or concatenate input videos implicitly. At most 10 videos are supported, within 30 seconds combined. Editing requires every video input to be 4–30 seconds, preserves the selected source's ratio and actual duration (including fractional seconds), and requires omitting `durationSeconds` and `aspectRatio`. If the user requests another duration or ratio, plan a separate trim or reframing operation. Extension preserves the source ratio; its selected 4–30 second duration is the new output clip's duration, not the source duration plus an increment. Reference generation supports independent output duration and ratio.\n- Seedance 2.5 supports `options.draft: true` for a 480p preview. Draft creation and final rendering always use KIE. Use Draft Mode only for the generation operation (text, image, or reference generation); source-locked editing and extension require standard mode. Quote the draft separately; do not claim that the initial credits include the final render. The user can render the same draft at 1080p within 7 days using the final-render icon on its Widemile card and confirming the displayed credits. Do not regenerate from the draft video as a reference to simulate final rendering.\n- For Kling 4, use only `kling-4.0/reference-to-video`, with optional image/video references and tokens `@image_1`/`@video_1`. Output: 3–30 integer seconds, 720p/1080p, auto/16:9/9:16/1:1/21:9. Files: 10 MB each; at most 5 videos, 2–30 seconds each and 30 seconds combined. Use `options.keepOriginalSound` (default false) for source video sound; omit `generateAudio`. Elements, voices, uploaded audio and frame mode are unavailable.\n- Generate audio for video by default, except Kling 4 original-sound retention as described above. For always-on-audio models, omit `options.generateAudio`. For configurable audio, set it to true unless the user explicitly requests silence; then select a compatible configurable model and disable it. Never describe an always-on-audio model as silent or audio-incapable.\n- Every video plan outside source-locked Seedance editing must declare an integer `durationSeconds` supported exactly by its selected product model. A final deliverable duration is an editing goal, not one generation: split longer work into independent clips and assemble them in Editor. `shotId`, when supplied, is only a concise optional workflow label and does not need to exist in the Brief.\n- Read the human Brief for creative intent, then select exact stable references from the live reference library and bind their purpose explicitly in the generation plan. Never parse reference ids, model ids, duration, or workflow state from Markdown, and never substitute unrelated or placeholder media. Keep generation prompts in English while preserving quoted dialogue, voice-over copy, lyrics, visible text, proper names, and other exact content in the requested target language.\n\n## Storyboard-driven video\n\nUse this branch when the user asks to develop video from a storyboard or the storyboard is useful to the current production. Read `prostudio_storyboard_get` for the latest relevant scenes. The storyboard remains optional for all other work.\n\n- **One scene = one video generation.** Prepare one independent plan and one complete video prompt for each requested scene. Its ordered inquadrature/shots are the cuts within that same clip. Use the stable scene UUID as the plan's optional `shotId` workflow label.\n- Write the prompt as an explicit ordered cut sequence. For each cut, carry over its action/composition caption, chosen visual reference, camera movement, focal length and timing when specified. Describe the transition into the next cut and preserve subject, setting and visual continuity. Preserve the user's choices and requested style.\n- Bind the selected shots' completed image `frameGenerationId` values through `continuityGenerationIds`, deduplicated in first-cut order. These are generation IDs, not reference-library IDs. Copy scene, shot and generation UUIDs exactly from the latest storyboard response; never shorten or reconstruct them. Before preparation, verify every cut-to-source mapping against the resolver output. If resolution fails, re-read the canonical storyboard and compare and recopy its exact source IDs before retrying, rather than repeating an unverified ID or changing the image. Resolve the full scene reference list successfully before using its tokens. Never merge token maps from individually or partially resolved subsets: token positions may change in the full list. If it remains unresolved, report the specific scene as blocked; do not present guessed bindings as a ready prompt. Other scenes can continue. Resolve them with `prostudio_references_resolve`, retain every selected image, and associate each cut with the actual binding token returned for that source ID. A reused image can guide several cuts with the same returned token. Add other references only when they serve the requested scene.\n- A cut without a selected image remains valid textual direction; do not force image or reference creation. Choose a model that supports the full scene's duration, selected references and intended cut sequence. Preserve each specified shot duration and give every cut a positive time interval in the prompt. The cut intervals must be ordered, contiguous, non-overlapping and cover the full scene. When every cut is timed, their sum is the scene duration. When some cuts are untimed, reserve positive time for each of them: the scene duration must exceed the sum of the specified durations, and added timings must be identified as estimates. Never assign zero remaining seconds to an untimed cut or shorten a specified cut silently. Choose a supported total that preserves the scene rhythm; if no exact compatible duration exists, report the limit before preparation. If the chosen model cannot accommodate the scene, explain the specific limit and agree on a compatible model or scene revision before paid preparation; do not silently omit images or split the scene into separate per-shot videos. Unrelated production work can continue.\n- Follow the normal prepare, exact-price consent and execution flow for each requested scene. When recording its completed video, use `prostudio_shot_generation_set` with `shotId` equal to the scene UUID and `generationId` equal to the completed video. This association stays in operational state; the storyboard shots' `frameGenerationId` fields continue to point to their chosen image references. Editing a storyboard alone never requests video generation.\n\n- Billing display is controlled by the authenticated account. When `billingDisplay.unit` is `EUR`, present the returned euro charge and balances; otherwise present credits. Preserve prepared execution arguments and `expectedCredits` exactly, without displaying those machine values as prices.\n"
}SHA-256 of public snapshot: 1fab1e15fb74e33dbcb478a3cdb14efcfcbe6fbf6ab2ffcb8e7a043fa405e89d