Skills and saved generators

Use a skill to influence how an authoring node writes. Use a saved generator to apply a reusable node setup, such as its model, format and prompt.

Reusable item	What it helps you reuse
Skill	Rules for a writing task or technique.
Saved generator	A ready node setup for a particular result.
Saved prompt	Text you want to use again.
Element	A character, object, location or style sample.
Design system	Saved graphics, component templates and motion for decorating a series.

The studio library calls these Skills, Generators, Prompts, Elements and design systems; in Russian, Скиллы, Генераторы, Промпты, Элементы and дизайн-системы.

Choose a skill for the writing step

Skills attach to authoring nodes, where text is written. Image and Video nodes execute the resulting prompt; they do not accept authoring skills themselves.

Through MCP, list_skills describes the available skills and which nodes they fit. Its optional node_type and model arguments help narrow the selection. Get those identifiers from the current studio catalog.

For example, a camera technique belongs in the instructions that produce the video prompt. The static-camera guide explains that technique and what to check in the result.

Reuse a visual series with Decoration

Decoration (Оформление) uses the original HyperFrames, Impeccable, Stitch and Anthropic design skill packs inside its authoring step; you do not need to attach them individually from Skills. V2 views actual timed frames and uses transcript, connected scenes and optional recommendations to make design decisions. Without a connected style, it creates an authored kit. Saved CSS, component templates and motion are reused physically by later episodes, rather than interpreted again from an MD description. You can review the result and save or export its style in Series design.

A Design system (Дизайн-система) is an optional way to manage the series’ saved style. Load the authored style created by Decoration from the library, or use Editorial, Signal or Kinetic with a Document (MD) as starting rules. Use ★ Save to library to keep an edition. Reusing that entry in another project, or carrying the nodes in a recipe, keeps the same complete saved profile and pinned skill revisions. A generator’s text recommendations can instead go to Decoration’s контекст input for the current episode.

The library also contains ten built-in design systems with full-film demonstrations. Preview them before adding one to the canvas, then connect it to Decoration. Previewing a demonstration starts no paid generation. See the design-system gallery and reuse workflow for the systems, saving your own edition and carrying it into another episode.

Creating a new composition is paid. V2 uses a fixed price based on the measured video duration, subject to the displayed minimum charge; analysis, authoring and the initial render are included. The older explicit directions retain their spending-ceiling and actual-token-usage pricing. Correcting text and rebuilding a saved composition do not call the author again; that rebuild remains $0. Follow subtitles and decoration for the current quote, connections, manual edits and server export.

Choose a generator for a repeatable setup

list_generators returns built-in setups and your own saved setups. Personal entries have mine: true. Use the returned description and node type to choose a suitable setup, such as a character design board.

Applying a generator configures the node; it does not generate its result. You can inspect the graph and estimate the next run before spending.

For MCP graph editing, apply_skill and apply_generator are operations inside apply_graph_ops, not separate top-level tools. They use IDs returned by the corresponding lists.

Rewrite a music brief with Music Caption Rewriter

Music Caption Rewriter is a separate node in Tools (Инструменты). It runs the original MiniMax Music 3 skill with its bundled reference library. Use this node directly; it is not another skill to attach through list_skills or apply_skill.

Its скил (skill) field picks the music model the prompt is written for. MiniMax Music 3, the default, is the original skill described here. ElevenLabs Music v2.5 is a studio skill written from ElevenLabs’ prompting guide, because ElevenLabs publishes API instructions rather than a prompt rewriter. It returns one short English paragraph. The paragraph covers genre, mood, instruments, tempo and key, production era, vocal delivery and the arrangement in order, without artist names or lyric lines. On ElevenLabs the lyrics travel inside the same prompt, which is limited to 4,100 characters. The node therefore keeps the paragraph within what the lyrics leave, at most 600 characters. It refuses before any charge when the lyrics leave less than 200 characters. A paragraph that breaks these checks gets one correction, then the run is refused. The pack version and update controls below apply only to the MiniMax skill.

Supply a Caption describing how the music should sound: genre, instruments, tempo, mood and vocal character. Put only the sung words and bracketed section/control tags, such as [Verse] and [Chorus], in optional Lyrics. Keep the song title, musical-style paragraph, Markdown formatting and assistant commentary out of Lyrics. Both fields accept typed text or a connected text input; a connected input replaces its matching field.

For a quick start without your own Caption, pick the жанр (genre), настроение (mood) and вокал (vocals) chips under the node, connect the lyrics and run. Each chip is optional and starts at auto. Genre lists the 18 genre families of the MiniMax skill; the skill still picks the reference templates within the family. Mood is calm, driving, epic, dark or upbeat. Vocals default to как в лирике (as in the lyrics), which keeps the voices from the lyrics breakdown, or override them with female, male, duet, choir or no vocals. The node turns the chosen chips into a short English style line in code, without another model call. With no Caption, that line becomes the Caption; with your own Caption, it is added as an explicit constraint that wins where the two conflict. Без вокала (no vocals) asks both skills for an instrumental, so the run is refused while Lyrics are connected. A run needs a Caption or at least one chosen chip; with all chips on auto, the request and cache are the same as before.

Lyrics may arrive in any form, for example from an ordinary text node: Markdown, performer names, stage directions and comments are all accepted, and the breakdown below removes them from the output.

Put the film’s script or scenario, narrative development, exclusions, duration and other context in дополнительные указания (additional instructions), or connect up to ten текст N inputs, which are included in order. For example, Caption can describe “warm acoustic folk, soft female vocal, no drums”, while an extra text input explains the scene and where dialogue needs space.

The node produces two text outputs:

• промпт: a new music caption with Global Metadata, Vocal Details and Arrangement sections. It is written in English unless you explicitly request another language.
• лирика: with the lyrics breakdown on, a clean sheet of section tags and the lines that are actually sung or spoken; otherwise the supplied Lyrics unchanged. The node does not write or reword lyrics. See below.

Connect промпт to Music’s промпт input. Connect лирика to Music’s matching input only when you want vocals and have selected a vocal-capable model. This caption format is designed for MiniMax Music 3. An instrumental request should remain instrumental; review the generated caption before making music. The skill uses lyric tags for section direction and is instructed not to reproduce the lyric lines in the caption.

The rewrite is a separate paid text operation. The displayed amount is a spending ceiling; successful runs settle against actual token usage. It does not include the later Music generation. If the skill cannot complete its required reference reads or produce a complete result within its limits, the node reports an error rather than substituting an ordinary rewrite. Oversized input is rejected without silently cutting text.

A separate model check compares the rewritten caption with the installed skill and your complete inputs. A grounded rule violation allows one repair and another check; an unavailable or invalid check, or a second rejection, stops the run with an error. These steps share the same spending ceiling. This check does not rewrite your Lyrics. A model check can miss a violation, so review the caption and lyrics before generating music.

Lyrics breakdown: clean sheet and voices

Nodes already run keep their earlier result and cache until you run them again.

When it is on, the node first splits your Lyrics with a separate studio skill, within the same spending ceiling and time limit. The result is a list of voices and a list of sections in time order, one voice per section; a dialogue turn is its own section. Each section has a delivery: sung, spoken, whispered, chant or instrumental. The лирика output is then assembled from section tags and lines only. Removed, with a reason each: the title, the style paragraph, separators, performer-name labels, stage directions, section headings and comments. A section tag already written as a clean English bracket tag is kept as written, including its timing, and an instrumental section stays without lines so the model does not sing over your dialogue.

Nothing is dropped silently. The card shows how many fragments were removed and why. Selecting the node shows the read-only разбор лирики field: voices by section and every removed fragment verbatim with its reason. Code checks the model’s answer: every word of the sheet must come from your text, and every word of your text must be either kept or listed as removed. One failed check allows one correction, then the run is refused with an error.

On Eleven Music 2.5 the voices are separated only with a composition plan. Music sends one when the промпт and лирика wires come from this node with the ElevenLabs skill, the sheet matches the text on the wire, and it has two or more voices or a section that is not sung. A chunk is made per section; code computes the lengths from the lines and scales them to Music’s seconds. The vendor allows 3 to 120 s per chunk, up to 30 chunks and 10 minutes in total; a sheet that does not fit is refused with the length needed, never trimmed. The price is the same as without a plan: per started minute of Music’s seconds. On MiniMax Music 3 the sheet is sent as lyrics tags with the voice map in the caption, and it sings in one voice. A duet is not guaranteed, and the Music card warns about it. In one listening test only the composition plan produced distinct voices; this is a single observation, not a guarantee.

An older caption remains visible for inspection. When you explicitly run its node again, an unchecked result is not reused as a checked result; the run must obtain a caption that passes the current check before continuing to Music.

Keep a node on a chosen skill version

The node settings show its source revision and a link to the original skill. Each node keeps its selected version. An administrator can use Check GitHub for updates and Update this node to install a complete revision and select it for that node. Other nodes keep their versions; a failed installation leaves the installed version intact. Use installed version lets a node select a revision already installed by an administrator. Updating a node changes the version used by its next rewrite.

This node currently runs through the Studio interface, not MCP run_node. These input and update contracts do not establish caption quality or musical quality. See Music models, sound design and the node reference.

Keep the choice tied to the task

Start with the result you need, then choose the authoring node and an applicable skill or generator. A model-specific instruction should match the model that will receive the prompt. Use the model guide and live catalog to check that choice.

What the built-in skills are for

The current built-in library contains a base skill, model adapters and craft skills. Availability in a particular node still follows the skill’s scope and your account. Use list_skills for the IDs and compatibility rather than copying a name into an arbitrary field.

Skill	Role and when to use it	What to review
baza-studii	Base craft guidance, preset on Prompt Scriptwriter; use for action that fits the scene’s duration.	Concrete action, readable staging and an explicit sound intention. You can remove and reapply it.
kling-v3-shot	Adapter for Kling v3 prompt writing.	Scene, dialogue and sound instructions fit the selected Kling operation.
seedance-2-shot	Adapter for the Seedance 2.0 family.	Reference roles and the ordering of action match the actual supplied inputs.
seedance-2-5-shot	Adapter for Seedance 2.5.	Generation, editing and extension are not confused; references have specific jobs.
grok-video-shot	Adapter for Grok Video prompt writing.	The main action and sound intention remain clear without unnecessary scene repetition.
gemini-omni-shot	Adapter for Gemini Omni Flash.	The instruction matches generation or editing and states the intended sound.
happyhorse-shot	Adapter for HappyHorse.	Camera, action and any editing request remain focused.
minimax-h3-shot	Adapter for MiniMax H3.	Reference use, dialogue and camera intent are clear for that route.
flux-3-shot	Adapter for FLUX 3.	The prompt agrees with its frame-based inputs and selected audio setting.
wan-shot	Adapter for Wan.	Longer action is organized into coherent shots rather than leaving gaps.
yazyk-kamery	Craft guidance for precise framing, camera movement and joins between shots.	One clear camera intention and readable transitions.
svet-operatora	Craft guidance when lighting, time of day or atmosphere carries the scene.	A motivated light source and contrast that support the subject.
tishina-v-kadre	Craft guidance for a scene with no spoken lines.	No accidental speech cues; listen to the resulting clip.
bez-vody	Craft guidance for replacing empty quality claims with concrete instructions.	The revised text still preserves all meaningful requested content.
zhivaya-postanovka	Craft guidance for active, naturally staged scenes and reference analysis.	A visible event rather than subjects merely posing.

These are descriptions of purpose, not the protected instruction text. An adapter is not a guarantee of model compliance. Check the model guide and technique-specific evidence before treating advice as a measured result.

Prompt for the model uses applicable craft guidance and the target model’s handling; do not attach every model adapter to it. Image and Video execute prompts rather than accepting authoring skills. Some older metadata mentions the legacy Director; that does not make it a current node to add.

Choose a built-in generator

The table describes the current 22 selectable configurations. Image configurations produce images when run; text configurations produce text. Two configurations open Prompt for the model with the appropriate task. Always inspect the applied model, format and estimate before generation.

Configuration	Supply	Result and next use
Character Design Board / Лист персонажа	A chosen character image and brief.	A multi-view character sheet; check identity across views, then keep it in an Element.
Location Design Board / Лист локации	A chosen location image.	Views, light and material details of the location; check that it retained the supplied place.
Object Design Board / Лист объекта	A chosen object image.	Views and details of a prop, vehicle or object; inspect shape and material consistency.
Hero face / Лицо героя	A character sheet.	A clean head-and-shoulders image; check that it is the same character.
Hero figure / Герой в рост	A character sheet.	A clean full-length character image; inspect proportions and clothing.
Clean object / Предмет чисто	An object sheet.	One whole object without panels; inspect its outline before using it as a reference.
Storyboard Sheet / Лист раскадровки	The complete script.	A visual overview across scenes; check scene order and labels.
Scene Storyboard / Раскадровка сцены	One scene and its Elements.	A numbered shot sheet for that scene; use only a video route that accepts the sheet.
Prompt for the model / Промпт под модель	Sample plans in context, your brief and Elements.	A model-specific video prompt; review the selected aspects before generating video.
Cast for the story / Герои под сюжет	Detailed sample plans.	Text for a new opening frame; inspect the new cast and setting before creating the image.
Clip by the story / Ролик по сюжету	Detailed sample plans plus your own subject inputs.	Prompt for the model with the plot preset; inspect events, timing and chosen aspects.
Ad idea · 30 s / Рекламная идея · 30 с	A product photo and any necessary product facts.	A commercial brief; verify factual claims and the product’s identity.
30 s director · Seedance 2.5 / Режиссёр 30 с · Seedance 2.5	The ad brief and product photo.	A timed video prompt; check it against the selected Seedance 2.5 route before rendering.
Social Profile Portrait / Профиль в соцсети	A person’s image you are entitled to use.	A stylized portrait concept; inspect identity, framing and any generated lettering.
Music Prompt / Промпт музыки	A script or musical brief.	Instrumental-music direction; check compatibility with the currently chosen music model.
Lyrics / Лирика	A script and intended song structure.	Only sung words and bracketed section/control tags, with timing intentions where relevant; review words and avoid collisions with dialogue.
Song Style / Стиль песни	The script and desired musical approach.	A musical style paragraph for Caption or Music’s prompt input; keep it separate from Lyrics.
Video Cover / Обложка ролика	A cover brief or the script and relevant references.	A cover image; check mobile readability and truthful representation of the film.
TikTok SEO	The finished film’s actual content.	Draft descriptions, keywords and related packaging; review claims before publishing.
Instagram SEO	The finished film’s actual content.	Draft captions and a packaging checklist; verify the selected platform’s current requirements.
YouTube SEO	The finished film’s actual content.	Draft titles, description and cover guidance; check that they match the film.
Facebook SEO	The finished film’s actual content.	Draft post/group text and packaging; choose the intended audience and review it.

Location Design Board has an automatically selected drawn variant; it is not an extra selectable library row. Social packaging generators draft text and guidance: they do not publish or guarantee reach. Their suggestions still need editorial review.

The shipped Lyrics generator asks for a clean sheet without a title, style paragraph or assistant offer. Existing nodes and personal saved generators keep their own instructions and prior results; they are not silently migrated or cleaned. To use the current preset, create a new node from the built-in Lyrics generator, then review its result before connecting it to Lyrics.

The “30 s director” configuration is a Text generator; it does not recreate the legacy Director node. A generator’s name expresses its intended task, not a promise that every currently selected model supports that output duration.

Scriptwriter reports and additional craft

Some studio deployments let Prompt Scriptwriter load additional planning craft while working. These internal capabilities are not extra library skills for you to attach individually. Inspect the writer’s report for what was loaded, its notice, findings and the critic’s outcome.

Availability and fallback are reported by the run. Do not infer that the advanced path ran merely because text was returned, or that a failed check means the text is safe to accept. The generated scenes remain yours to review before paying for media.

Follow a consistent character, video from a reference or graph editing to use these configurations in a complete workflow.

Catalog descriptions and Music Caption Rewriter contracts checked: 16 October 2026; the rewriter’s genre, mood and vocals chips on 22 October 2026. The visual-series workflow, gallery and V2 pricing distinction were checked on 22 October 2026. These are workflow and metadata checks, not generation-quality measurements or publication instructions.