> ## Documentation Index
> Fetch the complete documentation index at: https://mcp.shipcontent.com/llms.txt
> Use this file to discover all available pages before exploring further.

---
title: Make a Studio project
description: Make a video, carousel, image or Story at /{n}/studio, edit it with AI proposals, convert it, export each size and create posts for review.
section: workflows
order: 35
status: implemented
mode: both
reviewed: 2026-11-19
sources: [src/lib/server/owner/video-skills-access.ts, src/components/motion/use-export-size-choice.ts, src/components/motion/export-view-intents.ts, src/lib/mcp/agent-spending-budget-definitions.ts, db/accounts.sql, src/lib/media/motion-price-bound.ts, src/lib/server/media/motion-input-limit.ts, src/lib/server/media/MotionPlannerOpenAi.ts, src/lib/server/media/MotionSceneWriter.ts, db/billing.sql, src/lib/media/pricing-tiers.ts, src/app/[workspace]/(app)/studio/page.tsx, src/components/studio/studio-page.tsx, src/components/studio/tour/studio-tour.tsx, src/components/studio/tour/studio-tour-context.tsx, src/components/studio/tour/chapters.ts, src/components/studio/tour/progress.ts, src/components/guided-tour/guided-tour.tsx, src/components/motion/editor-shell.tsx, src/components/motion/add-panel.tsx, src/components/motion/stage.tsx, src/components/motion/strip.tsx, src/components/motion/inspector.tsx, src/components/motion/scene-panel.tsx, src/components/motion/layers-panel.tsx, src/components/motion/layers.tsx, src/components/motion/animation-panel.tsx, src/components/motion/audio-panel.tsx, src/components/motion/voiceover-panel.tsx, src/components/motion/captions-panel.tsx, src/components/motion/pause-panel.tsx, src/components/motion/loop-panel.tsx, src/components/motion/export-view.tsx, src/components/motion/export-view-controls.tsx, src/components/motion/export-mp4-preview.tsx, src/components/motion/brand-kit-panel.tsx, src/components/motion/template-menu.tsx, src/components/owner/jobs-menu.tsx, src/lib/media/motion.ts, src/lib/media/motion-templates.ts, src/lib/media/motion-brand.ts, src/lib/media/generation.ts, src/lib/media/types.ts, src/lib/media/audio-api.ts, src/lib/mcp/project-tools.ts, src/lib/server/media/project-agent-tools.ts, src/lib/server/media/media-http.ts, src/lib/server/media/MotionTemplates.ts, src/lib/server/media/MotionTranscripts.ts, src/lib/server/media/speech-audio.ts, src/lib/server/storage/MediaObjects.ts, src/app/api/motion/route.ts, src/app/api/motion/export/route.ts, src/app/api/motion/derive/route.ts, src/app/api/motion/revisions/route.ts, src/app/api/motion/links/route.ts, src/app/api/motion/files/route.ts, src/app/api/motion/hook/route.ts, src/app/api/motion/scene/route.ts, src/app/api/motion/transcribe/route.ts, src/app/api/motion/brand/route.ts, src/app/api/motion/templates/route.ts, src/lib/server/media/MotionAiRequests.ts, src/lib/server/media/MotionDocuments.ts, db/media.sql, src/lib/server/owner/owner-page.ts, src/lib/media/media-return-path.ts, src/lib/docs/handoffs.ts, src/components/studio/studio-home.tsx, src/components/studio/format-picker.tsx, src/components/studio/prompt-box.tsx, src/components/studio/credits-badge.tsx, src/components/motion/top-bar.tsx, src/components/motion/ask-ai.tsx, src/components/motion/proposal-panel.tsx, src/components/motion/sizes-view.tsx, src/lib/media/text-fit.ts, src/lib/media/export-checks.ts, src/components/motion/remake-drawer.tsx, src/components/motion/convert-dialog.tsx, src/lib/media/formats.ts, src/lib/media/convert.ts, src/lib/media/studio-view.ts, src/app/api/motion/text/route.ts, src/app/api/video-skills/route.ts, src/lib/media/video-skills.ts, src/lib/server/media/video-skills-http.ts, src/lib/server/media/VideoSkillCatalog.ts, src/lib/server/media/motion-plan-instructions.ts, src/lib/server/media/MotionStudioPlanHttp.ts, src/lib/server/media/clip-frames.ts, src/lib/server/media/MotionPlanner.ts, src/lib/server/media/media-agent-tools.ts, public/video-skills/kit/base.css, src/components/studio/skill-catalog.tsx, src/components/studio/skill-detail.tsx, src/components/studio/skill-setup.tsx, src/components/studio/skill-form.tsx, src/lib/media/video-skill-browse.ts, src/lib/media/video-skill-form.ts, src/components/studio/use-plan-start.ts, src/lib/media/video-skill-draft.ts, src/components/studio/use-skill-recovery.ts, src/components/studio/use-skill-assets.ts, src/components/studio/skill-recovery-notices.tsx, src/components/studio/skill-form-controls.tsx, src/app/api/credit-catalogue/route.ts, src/lib/server/media/CreditCatalogue.ts, src/lib/billing/credit-catalogue-api.ts, src/lib/media/credits.ts, src/components/owner/workspace-context.tsx, src/lib/marketing/post-sets.ts, src/components/motion/create-posts-panel.tsx, src/lib/media/video-workflow.ts, src/lib/mcp/agent-media-tools.ts, src/components/studio/workflow-detail.tsx, src/components/studio/skill-catalog-menus.tsx, src/components/studio/skill-catalog-results.tsx, src/components/studio/skill-card.tsx, src/lib/media/technique-index.ts, src/components/studio/technique-choice.tsx, src/lib/media/technique-choice.ts, src/lib/server/media/technique-choice-http.ts, src/lib/server/media/technique-choice-config.ts, src/lib/media/manual-video.ts, src/components/studio/manual-video-form.tsx, src/components/studio/manual-video-submit.ts, src/components/studio/studio-idea-context.tsx, src/components/motion/mobile-text-panel.tsx, src/components/motion/use-project-editor.ts, src/lib/server/owner/first-video-ui.ts, src/lib/server/skills/skill-run-access.ts, src/lib/skills/skill-recipe.ts, src/lib/server/media/video-skill-styles-config.ts, src/lib/server/media/MotionDocumentRows.ts, src/lib/server/media/project-agent-tool-writes.ts, src/components/motion/generated/context.ts, .env.example, src/lib/server/media/VideoSkillStyles.ts, src/lib/server/media/video-skill-styles.json, src/lib/media/video-style-spec.ts, src/lib/media/motion-style.ts, src/components/motion/style-tokens.ts, src/components/motion/generated/nodes-text-derive.ts, src/lib/server/media/MotionStudioPaidAiHttp.ts, src/app/api/skill-runs/route.ts, src/app/api/skill-runs/[id]/route.ts, src/app/api/skill-runs/[id]/answer/route.ts, src/app/api/skill-runs/[id]/cancel/route.ts, src/app/api/skill-run-files/[token]/route.ts, src/lib/server/skills/skill-run-http.ts, src/lib/server/skills/SkillRuns.ts, src/lib/skills/skill-runs.ts, src/lib/media/packet-scenes.ts, src/lib/media/packet-review.ts, src/lib/media/packet-budget.ts, src/lib/media/captions.ts, src/lib/server/media/PacketProduction.ts, src/lib/media/packet-music.ts, src/lib/server/media/packet-music-config.ts, src/lib/server/media/MediaJobDatabase.ts, src/lib/server/media/MediaCostEstimate.ts, src/lib/media/fal-opus.ts, src/lib/server/media/FalOpus.ts, src/lib/server/media/FalOpusPlanner.ts, src/lib/media/polished-plan.ts, src/lib/server/media/PacketReviewLoop.ts, src/lib/server/media/MotionStudioPlanFinish.ts, src/lib/server/media/PacketVisionPass.ts, src/lib/server/media/PacketAnalysisFrames.ts, src/lib/server/media/PacketRenderReview.ts, src/lib/server/media/PacketFramePublish.ts, src/lib/media/video-quality-rubric.ts, src/lib/media/video-self-verify.ts, src/lib/server/media/opus-planner-config.ts, src/components/motion/scene-motion-options.ts, src/lib/media/packet-drawn-scenes.ts, src/lib/media/scene-cut.ts, src/lib/media/quality-failure.ts, src/components/studio/project-fallback.tsx, src/lib/server/database/skill-run-reads.ts, content/video-skills/product-screen-reel/skill.json, experiments/twin/kit/sandbox/bundle-files.json, experiments/twin/kit/sandbox/kit-pin.json, src/lib/media/narration-fit.ts, src/lib/media/first-cut-workflow.ts, src/components/motion/use-packet-bed.ts, src/lib/billing/catalogue.json, src/components/motion/captions-layer.tsx, src/components/motion/motion-film.tsx, src/lib/media/caption-timing.ts, src/components/motion/seconds-input.ts, src/lib/media/studio-picker-assets.ts, src/lib/server/media/media-picker-identities.ts, src/lib/server/storage/media-object-render.ts, src/components/studio/reference-picker.tsx, src/lib/server/media/render-picker-identity.ts, db/migrations/20261109T1828_picker_render_identity.sql, src/lib/media/music-choice.ts, src/lib/server/media/elevenlabs-voices.ts, src/components/studio/skill-form-options.tsx, src/lib/media/motion-requests.ts, src/components/studio/use-studio-labels.ts, src/components/motion/voiceover-voice-picker.tsx, src/components/owner/elevenlabs-account-section.tsx, src/worker/AccountVoiceProvider.ts, src/lib/server/elevenlabs/ElevenLabsVoices.ts, src/app/api/motion/projects/route.ts, src/app/api/motion/projects/[id]/route.ts, src/lib/server/media/video-project-http.ts, src/lib/server/media/video-project-actions.ts, src/lib/server/owner/studio-project-access.ts, src/components/studio/project/project-page.tsx]
tools: [list_projects, get_project, create_project, edit_project, export_project, convert_project, prepare_seedance_reference, list_video_skills, get_video_skill, propose_generation, get_media_job, edit_post]
next: [/docs/reference/videos, /docs/workflows/media, /docs/contracts/revisions]
---

The **Images and clips** picker groups repeated exports of the same render. It keeps each stills slide separate and keeps a full export separate from a range clip. A saved selection follows the retained export. Existing exports with an unresolved render or slide keep their own entries.

A Studio project has one format. The owner picks it when the project starts. Each save is an immutable revision. The newest 100 revisions of a project stay, and so does every revision that a Seedance result names. The owner edits the project at `/{n}/studio?project=<id>`. An agent reads and edits it with the [project tools](/docs/reference/videos). A render is a worker job and calls no provider. Seedance shots, voiceovers, transcriptions and AI writing are paid provider calls. A text, link or thread post is not a Studio project; write it in Posts.

| Format | Limits | Sizes |
| --- | --- | --- |
| Video | 1 to 12 scenes, at most 30 seconds, 30 fps | 9:16, 1:1, 4:5 |
| Carousel | 2 to 10 slides | 4:5, 1:1, 3:4, 9:16 |
| Single image | 1 slide | 4:5, 1:1, 3:4, 9:16 |
| Story | 1 image slide, or a video of at most 30 seconds | 9:16 |

| Size | Pixels |
| --- | --- |
| 9:16 | 1080 × 1920 |
| 1:1 | 1080 × 1080 |
| 4:5 | 1080 × 1350 |
| 3:4 | 1080 × 1440 |

## Studio tour

On desktop screens of 1024 px or wider, the first ready visit to `/{n}/studio` offers **Take a tour** or **Not now**. Select **Studio tour** in the header to start or replay the chapter for the current view.

The guide explains grouped areas in Home, template or skill setup, Quick AI, Editor and Export. When video skills are on for the owner, the Home guide includes the skills on Home. The skill catalogue and skill detail pages do not have a chapter. Use the normal Studio controls to open another view. **Finish** marks the current chapter variant complete; **Skip tour** and **Close** end it without completion. Tour controls do not change content, generate media, render, download files or create posts.

When first-video controls are enabled, **Finish** closes the tour into the current Studio view. **Studio tour** still opens the chapter chooser for a later replay.

The browser stores the invitation choice and completed chapters for each signed-in owner, across that owner's workspaces. Progress stays in that browser; an active step is not saved. Close an open dialog or finish a pending action before starting a tour.

Updated 2026-10-16: the plan price covers planning only. A requested voiceover remains unfinished until the owner approves its exact saved-script quote in the editor. Captions have a separate approval based on the saved audio duration. Opening or reopening the editor never starts either paid child. The editor stays usable while an approval is pending. If a running voiceover returns to approval, the editor shows its refreshed quote and waits for another click. When billing does not require a quote, the owner can still approve. Updated 2026-11-05: on the **Make first cut** form, the badge charges planning only. The credit list under it adds length-based voiceover and caption estimates as later steps, and the catalogue credits per block for export (`media.render`). The credits per block can differ from the charge for one export.

## Start a project

Updated 2026-10-26: first-video controls are off by default and appear only for accounts enabled by the operator. When enabled, **Make a text video manually** works beside available or unavailable planning. It makes native text scenes and does not apply a paid style or call a planning, voice, music or footage provider. Existing access, storage, quota and render-price checks still apply.

1. Open the manual form. Set **Project name** (up to 80 characters; default **My video**), then the separate **Text shown in the video** (required, up to 80 characters). Add optional **Supporting text** (160) and **Closing text** (80). The project name is not shown in the video.
2. Choose the video size and durations. The opening defaults to 4 seconds and the optional closing to 3 seconds. Each scene can use half-second steps, from 0.5 seconds, within the total 30-second budget. Changing one duration does not shorten another scene or remove words.
3. Select **Create editable video**, edit the saved scene text, then export.

The general idea stays in memory while this owner and workspace keep Studio open, including style browsing and Back. **Clear idea**, successful use, leaving Studio, sign-out or a scope change clears it. A page read for another owner or workspace retires the previous Studio memory and uses that owner's presentation settings, even when the workspace number is unchanged. Reload and a new tab start empty. The idea is a reference; **Use idea as opening text** copies it only when it fits the 80-character limit. It is never shortened or copied into a paid skill automatically. Paid skill answers and paid-request recovery remain separate.

The manual form stays in Studio memory through Back or dismissal, including before submission. One manual create attempt also stays if its response is missing. **Check saved video** replays that exact request and composition. If that check is refused, restore access and check the same attempt again; a refusal does not prove the original creation failed. A second creation is disabled while the attempt is unresolved. **Stay** keeps tracking; **Stop tracking and leave** permits departure but does not cancel creation. Tracking ends on actual teardown. Reload, sign-out, tab closure and other navigation can lose it. If an earlier creation had no result, check Recent work before creating again.

Open `/{n}/studio`. Select a format: **Video**, **Carousel**, **Single image** or **Story**. For a Story, select its media: **Image** or **Video**. Then do one of these steps:

- Write a brief in the brief box on the Studio page, add images or clips if you want, and select **Make first cut**. **Search references** and **Next references** find older files in the loaded Library. Select up to 6 references. The server sends the brief, the saved business context and the media to OpenAI. This is a paid model call. The plan has 3 to 5 scenes for a video.
- When video skills are on for the owner, under **Start from a skill**, pick a [video skill](#video-skills), or select the **All … skills** link for the full catalogue. Only the 41 existing styles start a video: the skill asks for its own inputs in place of a brief, and **Start a video** is a paid model call. A workflow is non-runnable staged guidance. The other entries are references; copy their instructions for an agent.
- Select **Start a blank video**, or under **Your templates** select a template you saved from the editor. This starts a project with no AI call.

If the reply of **Make first cut** is lost, select it again. The browser sends the same saved request, and the server returns the saved draft; it does not send a second paid request. A plan that has no outcome after 150 seconds shows the status `unknown`.

If access ends after the server accepts a first cut, follow the recovery guidance for [`MOTION_PLAN_UNKNOWN`](/docs/contracts/errors#video-errors). The browser keeps the original request and its audio IDs, even if the project already opened. Use **Check the earlier request** to read the project or **Resume original request** to replay it. A failed read keeps the request. A read that finds no project releases the request when the request is 60 seconds old or older. A replay uses the same request and does not plan again. A confirmed matching `ready` or `failed` project releases the saved request. A definitive refusal before acceptance clears only the browser submission that still owns the saved request. A delayed refusal from another tab cannot erase a newer submission of the same request. Saved recipe answers, notes and options stay. If this browser cannot safely save and coordinate the request, it does not submit paid planning. Check browser storage settings or use another browser.

Updated 2026-10-23: each Motion AI model has its own price. A planning, hook, text or scene request uses `ai.motion.<kind>.<model>`, for example `ai.motion.plan.gpt-6-sol`. Studio sends `gpt-6-sol`. A remake of a clip with a reference video includes the seconds of that video in its estimate. Each price is the most that one request can cost. The server refuses a Motion AI request with more than 64 KiB of text (instructions, input and schema) before it sends it, and the hold is released. Shorten the brief or the business details, then try again.

Updated 2026-11-12: the `gpt-6-sol` tier calls OpenAI `gpt-6.1-sol`. The tier name and its prices do not change. Astra and Luna call `gpt-6-astra` and `gpt-6-luna`.

Updated 2026-11-12: the editor shows the first-cut voiceover script above the approve button. The owner can change it until the voiceover is approved. The change keeps the voiceover request ID, so a voiceover job already sent with the old script refuses the new one with `IDEMPOTENCY_CONFLICT` and is not paid twice. After the voiceover is done, a change fails with `MOTION_CHANGED`. Agents cannot change the script.

Updated 2026-11-12: a project keeps the plan that the model returned, as JSON: title, script and every scene. Open **Video settings** and expand **AI plan** to read or copy it. Script edits and scene changes do not change this copy. Projects planned before this date, blank projects and converted projects have no AI plan.

Updated 2026-11-13: the editor has no **Leave Studio** button. Use the Studio back button or the sidebar. A project that you open from a post still shows **Return to post**.

Updated 2026-11-14: the planning request that `MotionPlannerOpenAi` sends (planning, hook, text and caption calls) and the scene request that `MotionSceneWriter` sends pin `service_tier: "default"`, the standard service tier. They do not inherit a project Fast/auto tier. They also send `prompt_cache_options: { mode: "explicit" }` and mark no cache breakpoint. OpenAI documents that such a request writes no prompt cache, so its input does not pay the cache-write rate (1.25 times the input rate). By those documents, its token rates are the published OpenAI standard rates in `motionTokenPrices` in `src/lib/media/motion-price-bound.ts`: $10 input and $50 output per 1M tokens for `gpt-6-astra`, $2 and $10 for `gpt-6-sol`, and $0.10 and $0.50 for `gpt-6-luna`. This applies to a project without regional processing. No provider bill was measured. The credit prices do not change. This corrects the 2026-10-23 statement: before this change, a request could inherit a project tier or pay a cache write, so its price was the most that one request can cost only at the standard tier and with no cache write.

Planning, hook, text and scene requests carry a top-level `quote` separate from the creative input. Each has quantity 1. The browser keeps the request ID and its displayed quote through uncertain responses. In enforce mode, `PRICE_CHANGED` refreshes the displayed price and requires another owner action. Existing submitted requests retain their original mode and price. A successful provider response uses credits even when its content cannot be used. Unknown outcomes do not automatically retry or refund. OpenAI `incomplete`, pending, missing and unrecognized response states remain unknown with their available receipt and token evidence; HTTP 200 alone does not prove completion.

New projects use the workspace brand kit. The page calls owner routes under `/api/motion/`. These routes need the signed-in browser session and a CSRF token. An agent token cannot call them.

Credits in Studio show the relative size of an AI action. They are never shown in dollars. See [Export](#export) for render pricing. Workflow cost estimates are separate; see [video skills](#video-skills).

Studio, Quick AI and the editor panels calculate credits from the prices that the operator applied. When a price changes, the new estimate shows without a new deploy. The page reads the prices again when a panel opens, when the window gets focus, and at `validUntil`. Each action costs whole credits, and the costs of the actions are added. When the prices cannot be read, or an action has no price, the panel shows **Credits unavailable**.

The owner request `GET /api/credit-catalogue?workspaceId=<uuid>` returns the credit prices that the operator applied. It needs the signed-in browser session and current access to that workspace. An agent token cannot call it. The response has `Cache-Control: private, no-store`, so HTTP caches do not store it. The page keeps the latest snapshot in memory until its next read.

- `costs` has one entry for each action key, with `unit`, `creditsPerBlock`, `blockSize`, `version` and `effectiveFrom`. One action costs `ceil(quantity * creditsPerBlock / blockSize)` credits. Round each action up to a whole credit, then add the costs. For example, a 22-second transcription at 4 credits per 300 seconds costs 1 credit. An action that is not in `costs` has no price.
- `asOf` is the time of the read. `validUntil` is the time of the next scheduled price change, or null. Read the prices again at that time.
- `priceVersion` changes when a current price changes. Compare it; do not parse it.

### Video skills

Updated 2026-11-19: the skill directory is open to every owner as read-only. `/{n}/studio?skills` and `?skill=<id>`, `GET /api/video-skills`, `list_video_skills` and `get_video_skill` do not need a list. Starting a skill in Studio, `mode=techniques`, `?techniques=1` and skill runs also need `SHIPCONTENT_STUDIO_VIDEO_USERS`. For an owner who is not on that list, `engineRuns` is false, a skill page has no **Start**, and **Copy** gives the skill instructions only.

Updated 2026-11-09: the operator controls video skills with the server-only `SHIPCONTENT_VIDEO_SKILLS_USERS` setting on the app server. It is a comma-separated list of owner UUIDs. `*` allows every owner. Unset, empty or a list with an invalid item allows no owner. The setting grants no other access. For an owner who is not on the list, Studio shows no skill shelf, and `/{n}/studio?skills=1`, `?skill=<id>`, `?new=video&skill=<id>` and `?techniques=1` are not found. `GET /api/video-skills` and `mode=techniques`, `list_video_skills` and `get_video_skill` return `NOT_FOUND`. A plan request with a `skill` returns `NOT_FOUND` before any credit reservation. A saved or reserved plan with the same request ID still returns its saved result with no new charge. A saved project that started from a skill still opens at `?project=<id>`. A plain `?new=video` request is not changed.

**Find a technique**

When enabled in [local configuration](/docs/start/connect-local), **Skills → Find a technique** opens `/{n}/studio?techniques=1`. Enter a nonblank request of at most 500 Unicode code points and choose an optional tool. Search stores the complete request in `tq` and the tool in optional `tt`. These fields are separate from the catalogue search and filters. A bookmark with a valid `tq` runs that search; without a request, the form stays idle.

Search returns up to five technique and engine entries in the existing search function's default order. Tooling is the only filter; size, duration and runnable status do not change the result. Equal scores use ID order. Zero-score results remain, so an unrelated request can still return five alternatives. Titles and summaries keep their catalogue wording. **Inspect method** shows whole sentences from Shape, then Motion. The first result is **Recommended · Top search result**. Nothing is selected for you. Inspecting a preview or method does not select a result.

Select any result and choose **Continue with selection**, or choose **Accept recommendation** to open the first result even when another radio is checked. Both actions open existing skill detail. They do not generate media, create a job, call a paid provider or save a choice. References remain **Instructions only**. Changing the request or tool clears the choice, cancels the read and requires another search. Loading and edited results cannot continue. A failed read keeps the exact request for **Try again**. An empty result has no continuation; one result still requires an explicit choice. An unavailable detail shows the existing error without choosing another entry.

The free browser read is `GET /api/video-skills?workspaceId=<uuid>&mode=techniques&q=<request>&tooling=<tool>`. It requires the existing signed-in browser session, allowed browser origin and current workspace access; an agent token cannot call it. Only `workspaceId`, `mode`, `q` and optional `tooling` are accepted, once each. `q` must be well-formed, nonblank Unicode text within the request limit above. The original text reaches search without trimming, truncation or rewriting. `tooling` names one catalogue tool, not a comma-separated list. Unknown parameters, repeated parameters and list/detail fields are invalid. The response is `{ results, recommendedId }`: each result is an existing skill summary plus method text `how` and numeric `score`. `recommendedId` is the first result's ID, or `null` for an empty result. See [Error codes](/docs/contracts/errors) for refusal and retry guidance.

Detail navigation uses `skill=<id>&fromTechniques=1`, retains `tq`/`tt` and adds `techniqueFocus=<id>`. Detail **Back** and its header parent restore the submitted search and focus that result, or the chooser heading if it is absent. Focus never checks a radio. Chooser **Back** restores the catalogue context: optional `skills=1`, `q`, `kind`, `source`, `category`, `technique`, `tooling`, `page`, `cols` and `focus`. These fields keep their existing catalogue validators; page is 1–1000 and columns are 2, 3 or 4. Browser Back, reload, sign-in and session expiry retain valid submitted bookmarks. No selection is stored in a URL or persistent storage.

Chooser and chooser-origin detail handoffs share one encoding: spaces become `%20`, literal plus `%2B`, percent `%25` and ampersand `%26`. The catalogue context, including each leading `&`, is limited to 1536 canonical ASCII bytes. The complete encoded local pathname and query are limited to 8192 ASCII bytes. Both limits apply before a handoff is created and at repeated sign-in validation. An oversized context shows an explicit error to reduce catalogue filters; no field is truncated or dropped. Only valid numbered Studio chooser/detail tuples gain this allowance. Unknown or repeated keys, conflicting view selectors, arbitrary return URLs, fragments and selection parameters reject the whole return target. Invalid sign-in return targets fall back to bare Studio. Ordinary return targets keep their existing 300-character input limit and validators.

With the flag off, the entry is hidden and authorized direct chooser requests are not found. Existing catalogue and detail reads still work; a detail opened from a chooser returns to Skills. Sign-in return validation does not enable the feature. Existing catalogue search and written search-summary rules stay unchanged.

**Catalogue**

The catalogue has 1,119 entries: 475 separate references from the pinned Opus video collection, 601 donor references, the 41 existing styles and two workflows. No source records are merged. The donor references include 406 motion cards, 190 style, skill or technique entries and five engine references. The catalogue entries are shared. The read-time flags below depend on the caller's workspace.

A video skill has a `kind`: `style`, `technique`, `engine` or `workflow`. A style describes one short video. A technique or an engine entry is a reference. A workflow describes ordered stages.

`runnable` controls the Studio **Start a video** path. Publishing a recipe does not enable that path. `engineRuns` is true only when the entry has a valid published `recipe` and the caller's workspace is on the operator's skill-run trial list. A valid recipe has a `render` step and a `check-output` step, so it makes a video and checks it. The list, detail and **Find a technique** reads compute it the same way. The list does not grant access to another workspace. These flags do not start work. This release registers no `start_skill_run` tool. It publishes one recipe, for `product-screen-reel`, so that entry returns `engineRuns: true` only for a trial workspace; every other committed entry returns `engineRuns: false`.

Updated 2026-10-28: the project owner configures the trial list with the server-only `SHIPCONTENT_SKILL_RUNS_WORKSPACES` setting on the app server, for local or hosted use. Unset, empty or whitespace-only values allow no workspace. The exact value `all` allows every workspace; otherwise use a comma-separated list of workspace UUIDs. Spaces, letter case and empty list items are ignored. A read without a workspace ID stays off, including with `all`. Workspace owners cannot enable it through the catalogue.

For `product-film-and-reel`, `runsInstead` is `product-screen-reel` only when the caller's workspace is on that trial list and `product-screen-reel` is in the catalogue being read. It is null otherwise, and is null for every other entry. The committed catalogue contains `product-screen-reel` with a published recipe, so a trial workspace reads `runsInstead: "product-screen-reel"` for `product-film-and-reel`. The flag does not add a skill or start a run.

Since 2026-11-07, `musicBed` is true when a runnable style's Studio plan also buys a packet music bed: its packet names a bed, and `VIDEO_SKILL_STYLES_ENABLED` and `PACKET_MUSIC_ENABLED` are both exactly `true`. The catalogue, `list_video_skills` and `get_video_skill` reads compute it on each call; the catalogue file does not store it. It is false for every other entry.

Use the source filter `alchaincyf/huashu-art-motion` to browse its art styles and explainer techniques. Their English instructions adapt the donor's motion methods to the existing kit. The donor's music scripts, delivery checks and independent film-review process are not included. Its execution dependencies are not required to browse these references.

Use the source filter `lemomo-ai/lemo-opuscar` to browse 51 code-drawn film styles. `JohnHeibel/PDoomVideo` and `joeseesun/opus-video-prompts` have no licence at their pins. Their references are ShipContent's own descriptions of the ideas only: no donor text, code, prompt or picture is included.

The 41 existing styles have `runnable: true` and keep **Start a video**. A workflow has `kind: "workflow"` and `runnable: false`. It is not a reference. Every other entry is a reference with `runnable: false`. When `engineRuns` is false, references supply direction only: do not build, plan or run a video from them. The catalogue read does not run an engine or call a provider. Each entry has its own small preview image. `posterOrigin` identifies a retained composition render, a licensed source preview, or an original generated illustration. A fetched preview records its attribution URL, permission URL, licence and copyright. A generated illustration records its tool and prompt hash; it is not a source video frame. The Opus source posters are not copied because their separate reuse rights are not established. Full preview licence notices are supplied with the images. Each entry has `sourceId`. References also have `provenance` with their repository, pinned commit and licence. The original video entries also keep the author and original post link.

- `/{n}/studio?skills=1` is the catalogue. Select a category, or type in **Search skills**. **Kind**, **Technique**, **Tools** and **Source** filter the catalogue more. The URL keeps `kind`, `category`, `technique`, `tooling`, `source`, `q`, `page` and the column count `cols`, so a link opens the same view. A card link keeps the filters, the search and the page: setup **Back**, the detail **All skills** and the **Skills** level of the page header return to that view, and focus the selected card, or the **Skills** heading if the card is not in view. With no filter and no search, the catalogue shows one row for each category. A row shows its number of skills and the length range of its styles that start a video. Select **See all** to show all the skills of that category. A filtered view shows 24 skills on each page, with **Previous** and **Next**. On a phone, select the search icon to show **Search skills**, and **Filters** to select techniques, tools and sources. Each card shows the category or the length, and the number of paid services that the skill uses. Point at a still on the card of an existing style, or touch and hold it on a phone, to play that part of the sample. The keyboard focus on a card plays its first still. The sample does not play when the system setting for reduced motion is on. A reference card shows its preview still, which does not play, and **Instructions only** in place of the length and the paid services. Reference covers fill the card width with a proportional crop. The detail page shows the full square image. Playable sample stills keep their full frame. A workflow card says **Workflow**. A filtered view shows the styles that start a video under **Starts a video**, workflows under **Workflow**, and the references after them under **Instructions only**.
- `/{n}/studio?skill=<skill id>` shows one skill: its sample, its inputs, its length, its sizes with the platforms that can publish a video of each size, its tooling, its source credits and the credits of the third-party work in its kit. For a runnable style, **Play sample** plays the sample with no sound. Select a still next to the player to play the sample from that moment. Drag the bar under the player, or use the arrow keys on it, to go to a different time. The page shows the plan credits, an estimate that adds a voiceover and captions at the longest length of the skill, and the paid services of the skill. When `musicBed` is true, the plan credits include the 60 credits of the 30-second bed, and the paid services include **Music**. Voiceover and captions require separate approval after their exact inputs exist. Under **Give it to an agent**, **Copy instructions** copies the authoring handoff of the skill for a connected agent. Runnable SaaS/product-screen and concept-flow cuts require native editable output before richer renderer Build steps. An audio input shows **Only a connected agent can use this file**. A file input that takes more than one file shows its limit, for example **Image, up to 3**. A reference shows its preview still, its repository and pinned commit, and **Copy instructions** in place of **Start a video**. It shows no length and no paid services. A workflow shows its stages, owner stops and cost ranges, and no **Start a video**. Its preview is a still, with no sample player or agent handoff control. Under the still, a fetched preview names its licence, and a generated preview names Codex or Grok Imagine. **Tools and credits** shows the licence and copyright of a fetched preview. References keep guidance only. Existing skill IDs and URLs still resolve.

**Workflow entries** show stages, owner stops, paid steps and separate provider and agent-token cost ranges in USD. Use the Kind filter, `GET /api/video-skills?workspaceId=<uuid>&kind=workflow`, or `list_video_skills` with `kind: ["workflow"]`. A workflow can embed any number of technique skills. Existing style entries keep the 30-second limit and Studio sizes. Workflows can describe up to 180 seconds, 16:9 and optional owner-presenter tooling. These capabilities do not extend the Studio planner or renderer.

The **Product film and reel** workflow is non-runnable staged guidance for the owner's agent. It targets a 30–45 second 9:16 reel first and an optional 90–180 second 16:9 film. These guide targets do not extend Studio's formats or limits or install a workflow runner. Ask a connected agent to call `get_video_skill` with `id: "product-film-and-reel"` and follow the returned instructions for work outside Studio's supported cuts. A workflow ID is rejected by the Studio plan endpoint before a paid planner call. The detail response includes six agent-only technique skills under `workflow.techniqueSkills`. They do not appear as separate cards. `kitUrl` continues to serve the style browser kit.
The **Product screen reel** workflow makes one silent 9:16 reel, 15 to 30 seconds, from four to six screenshots, a headline and optional captions. Ask a connected agent to call `get_video_skill` with `id: "product-screen-reel"`. `runnable` is false. `engineRuns` is true only for a workspace on the skill-run trial list. Studio does not start this workflow. There is no paid call.
- `/{n}/studio?new=video&skill=<skill id>` is the start form of an existing style. It has one field for each input of the skill. Studio fills in the fields that the workspace already knows: the name, the colors and the logo of the brand kit, the website, and the description, the audience and the call to action of the business profile. A filled field has the mark **From your brand** until you change it. Studio does not write over an answer that you gave. Updated 2026-11-08: when video planning is unavailable on this form, the notice appears above the inputs. Select the size, the length (**Auto** uses the length range of the skill), and a voiceover and captions if you want them, then select **Start a video**. A submitted plan is a paid model call with the existing approval rules. When `musicBed` is true, the form shows **Background music**, on by default. While it is on, the form shows a music mood, the credits include the 60 credits of the bed, and the form says **Plan and a music bed. Voiceover and captions need separate approval.** Turn **Background music** off and the mood control is hidden. The form then says **Plan only. Voiceover and captions need separate approval.** The plan request sets `music.enabled` to false, so this video skips the bed and its 60 credits. The request sends `music.mood`. **Auto from style** keeps the style prompt. Any other mood shapes the bed prompt. The server still buys the bed only when `PACKET_MUSIC_ENABLED` is exactly `true`. A request that omits `music` keeps the earlier behavior. A workflow shows no form. The plan request refuses it before paid planning. A reference entry (`runnable: false` and not a workflow) shows no form, only **Copy instructions**, and the plan request refuses it with `INVALID_INPUT`.

The runnable setup keeps unsent answers, explicit clears, notes and options in this browser, bound to the owner, workspace and recipe content. Back and reload restore them before brand fill. Other browsers and agents cannot read this local draft. A changed recipe shows the old inputs as read-only reference data. Corrupt or mismatched saved data shows a read error instead. **Start fresh** replaces changed or invalid inputs only when no request is pending. For an uncertain submitted request, use the [saved-request actions](#start-a-project). A matching saved recipe can explicitly resume the same request. With a changed recipe and an uncertain request, only the saved-project read and **Back** remain available. Leaving setup cancels that read and prevents a delayed result from reopening the project. Browser storage failures keep current inputs on screen and prevent an unsaved paid submission. **Read saved inputs again** retries a failed initial read before editing; **Save inputs again** retries a failed write. Selected files outside the recent Library list are checked through existing owner pagination. An incomplete or failed check keeps the selection and blocks Start; **Check Library again** retries a failed check. A confirmed missing or wrong-kind file needs explicit replacement or removal.

For SaaS/product-screen and concept-flow cuts within Studio's supported formats and limits, native editable output is mandatory and takes precedence over renderer Build steps. Flattened footage does not replace it. External native authoring uses `create_project` and `edit_project`: MotionComposition version 3, SceneDefinition version 1, declared text/number/color-role fields, 30 fps, at most 30 seconds and 12 scenes. Save with `expectedRevisionId` and return the protected Studio project link. Keep a local sidecar with producer run, source revision, skill, kit and output identities. This adds no origin API. Rich renderer output is footage; baked graphics are not editable fields. Copy does not dispatch a CLI, sandbox or job. Browser **Start a video** continues to use the server planner. Both paths converge in the current editor, immutable revisions and existing export approval.

The browser examines the answers before it sends the plan request. A required text or link field with no answer shows **Fill this in.** A required image or clip field with no file shows **Choose a file.** In both cases, no request is sent.

The catalogue has two owner requests. They need the signed-in browser session. An agent token cannot call them.

- `GET /api/video-skills?workspaceId=<uuid>` returns `skills`, `total`, `matched`, `nextOffset`, `facets` and `source`. It returns at most 48 summaries. Set `limit` from 1 to 48 and use `nextOffset` as the next `offset` with the same filters. A null `nextOffset` means the list is complete. `matched` counts the filtered result; `total` counts the full catalogue. Use `q` for a search of at most 80 characters. Each entry of `skills` is a summary with up to three stills and `posterOrigin`. The optional `kind`, `category`, `technique`, `tooling` and `source` each take a comma-separated list. `kind` is `style`, `technique`, `engine` or `workflow`. A skill is in the result when it has one or more of the values of each given filter. An unknown filter name or kind, category, technique or tooling value returns `INVALID_INPUT`. An unknown source returns an empty result. Source IDs are `opus`, `shipcontent` or a donor repository name.
- `GET /api/video-skills?workspaceId=<uuid>&id=<skill id>` returns `skill` and `source`. The skill has its instructions, inputs, samples, `posterOrigin` and `sources`. A workflow also has the `workflow` details described above. An unknown ID returns `NOT_FOUND`.

An agent reads the same catalogue with the [video skill tools](/docs/reference/videos) `list_video_skills` and `get_video_skill`. They take the same filters and IDs, give full links to the stills and call no provider. `get_video_skill` also gives `kitUrl`.

For runnable skills that need richer footage, the renderer Build steps use HyperFrames with a shared kit at `/video-skills/kit/`; the native-output requirement above takes precedence for SaaS/product-screen and concept-flow cuts. Keep each entry's `credits` with videos that use the kit files they name. The kit has `base.css` and helper scripts for timing, text, drawing, canvas, pixels, particles, physics, shaders, 3D, print effects and animated icons. GSAP 3.13.0, three.js 0.186.1 and the fonts are not in the kit; load them from their own packages. The 18 animated icons come from icons-animated: the lines are Hugeicons free icons and the motion adapts lucide-animated. All three are under the MIT licence. The `animated-icons` skill gives them in `credits`, and `icons/LICENSE` and `icons/THIRD_PARTY_NOTICES.md` stay with each copy.

To make the first cut from a skill, send the optional `skill` field on the plan request `POST /api/motion`: `{ "id": "<skill id>", "values": { "<input id>": "<text>" } }`. The value of an image or video input is a list of Library file IDs, and each ID must also be in `referenceIds`. With `skill`, `brief` is optional and `format` must be `video`. An audio input of a skill is not part of the plan request.

The server examines the answers before the paid model call. It returns `INVALID_INPUT` when:

- the skill does not exist, or `values` names an input that the skill does not have
- a required input has no value, or a value has the wrong type or too many files
- a file is not of the kind of its input
- `size` is not one of the sizes of the skill, or `seconds` is outside the length range of the skill

With no `seconds`, the plan uses the length range of the skill. The planner gets the skill title, description and scene direction. It gets the answers with their input labels as data, separate from the brief.

Updated 2026-10-30: Each of the 41 existing styles also has a typed style: pacing, beats, shot types, typography, motion, palette, narration tone and audio. The style is off until the operator sets `VIDEO_SKILL_STYLES_ENABLED=true` on the app server. Only the exact string `true` enables it. When it is on, the plan request uses the style, and the saved video keeps it as `composition.styleSpec`. Since 2026-11-12, a style can be added or changed only for an owner listed in `SHIPCONTENT_VIDEO_SKILLS_USERS`: a Studio save, a Studio create or convert, and an agent `create_project` or `edit_project` for an owner who is not listed behave as when the style is off, so such a write that adds a `styleSpec` or changes the saved one returns `INVALID_INPUT`. A write can keep or remove a saved style for any owner. The style has these effects:

- **Applied:** the number of scenes, the scene durations and the scene motions limit the plan schema. After the plan, Studio sets the text position of each scene that is not footage, the caption style and position, and a fixed palette as the brand colors. The renderer sets the headline weight and case, also in animation scenes, the text enter and exit frames, and a dark or brand background. In a styled video, foreground colors adjust for contrast on scene, card and accent fills. A text layer of an animation that sets its own weight or upper case keeps it.
- **Planner fact:** the beats and their share of the length, the shot type of each cut, the headline length, the words per second, the voiceover tone and the full palette (source, colors, surface and grade). The planner gets them as data. The shot types are `null` for each existing style, because their sources have no shot data. A style copied from a video can fill them. The renderer does not reframe a shot or apply a color grade. Nothing checks the model's use of these facts. The voice provider gets only the words and the voice, so the tone does not change how the voice speaks.
- **Stored only:** the font, the easing curve and the tempo. Studio keeps them on the video and does not apply them. There is one exception since 2026-11-01: when the font note of a packet style opens with a serif, as Concept explainer and Article to video do, its headlines use the bundled serif at weight 400.

Concept explainer, Generated footage film and Kinetic manifesto also have a packet. Since 2026-11-10, Article to video and Hype edit also have a packet. The packet owns the scene template. The planner fills the words, and for Generated footage film it fills the picture subject. Concept explainer draws a diagram and buys no still. The Generated footage film packet asks for one FLUX Schnell image job for each scene, with no placeholder picture, but Studio does not start those jobs. Since 2026-10-31, when `VIDEO_SKILL_STYLES_ENABLED` is exactly `true`, the Studio plan request refuses Generated footage film with `INVALID_INPUT` before it reserves credits or plans, because Studio cannot attach the bought pictures to scenes yet. When the flag is off, no skill has a style or a packet, so Generated footage film plans and reserves like any skill with no style and buys no packet picture. Kinetic manifesto uses kinetic type and buys no picture. Article to video draws a diagram like Concept explainer and buys no still. Each Article to video scene is 4 to 6 seconds. A shorter scene fails with `MOTION_PLAN_FAILED` before the plan is saved. Hype edit uses kinetic type like Kinetic manifesto and buys no picture; its packet scenes do not show the clips or photos that the start form accepts. Both name a music bed. Since 2026-11-10 the planner of a packet style treats a link in the skill answers as a reference only: it cannot open the link, so it takes the facts from the brief and the written answers. Since 2026-10-31 the first scene of Concept explainer and Kinetic manifesto shows its headline and main parts from frame 0, with no entrance. Parts inside a diagram block still enter in steps. Since 2026-11-01 every Concept explainer scene shows its headline, step line, body, caption and progress marks at the cut, and the first scene also has two of its tokens in place. Since 2026-11-04 the opening headline still keeps its entrance: its words are legible at frame 0 and rise into place, and the headline of a later Concept explainer scene enters word by word after the cut. The free kinetic-text check does not count a kinetic-text node that is drawn and has no entrance. Each Kinetic manifesto scene also has an accent shape that pops in at 1.6 seconds, so no scene holds still for more than 2.5 seconds. In a packet video, the outgoing generated scene holds its drawn frame up to the cut. The incoming scene starts part way into its move and settles within 8 frames. A packet can name a music bed. The plan records that bed and ducks it when an audio file is attached. Since 2026-11-06, the server flag `PACKET_MUSIC_ENABLED` controls the bed. Only the exact string `true` enables it, and it is off by default. With the flag off, the plan records the bed and does not buy it. With the flag on, the owner's plan enqueues the bed as one `text-to-music` media job on `fal-ai/elevenlabs/music`. The job asks for an instrumental bed in whole seconds, from 3 to 600. Since 2026-11-07 the bed is 30 seconds, the video limit, so it still covers a video that the narration fit makes longer. The media worker submits it once, and stores the audio like any other generation output. fal bills each started output minute, so the plan reserves $0.80 for each started minute against the five-dollar cap. A refused or failed enqueue keeps the plan and counts that reserve as failed. Since 2026-11-07 the bed has the price `media.music`: 60 credits for each started minute, at 1 credit for each $0.01 of the fal price. The database refuses a bed under 3 or over 600 seconds with `INVALID_INPUT` before any provider call. The plan sends no quote for the bed, so when billing enforcement is on, the job is refused with `PRICE_CHANGED`. When the recorded bed job finishes, the editor adds the bed once, as one music track over the whole video, and ducks it under the voiceover. An owner who removes that track keeps it removed. The editor adds only the job that the plan recorded. Since 2026-11-07, when the first-cut voiceover of a packet video finishes, the editor fits the scenes to the saved narration length. The video then lasts as long as the voice. A diagram or a still scene lasts 4 to 6 seconds when that is possible. Since 2026-11-10 a kinetic-type scene keeps the scene length range of its style when that is possible. That range is at least 2 seconds and at most 6 seconds. A short narration keeps the shortest scenes, and the bed plays on after the voice. A narration over 30 seconds stops at 30 seconds. The bed covers the new length. A narration with no saved length keeps the planned scenes. A video with no packet is unchanged. Studio and agents cannot ask for a bed: the generation route and `propose_generation` refuse `text-to-music`. Customer resale rights for this music are not established, and the plan record says so. A plan stores a review record of sampled frames. The text planner can send the brief to Claude Opus 5.5 on fal, through `openrouter/router`. Since 2026-10-31 this route is off unless `MOTION_OPUS_PLANNER_ENABLED` is exactly `true`; with it off, every plan uses the OpenAI planner, even when the server has a fal key. Before that send it quotes a ceiling of two times the Anthropic list price: $4 per million input tokens and $20 per million output tokens, with an output limit of 4,000 tokens since 2026-11-08 (1,600 before). The request sets reasoning on. Reasoning tokens count inside that limit. A plan reply that uses the whole limit and does not parse as a plan fails with `MOTION_PLAN_FAILED` and a message that says the reply was cut off. That message does not ask for a new brief. A complete plan at the limit is used. A scene revision and a vision call keep their 1,600-token limit. The prompt includes the brief, the brand and the skill answers, and it asks for one JSON object. The fal route has no JSON schema field. The prompt names the required output keys and one valid example. The prompt cap is 8,000 UTF-8 bytes. A multi-byte character counts as its bytes. The two-times factor is an assumption for router markup. It is not a fal tariff. The plan reserves that ceiling, plus the narration grain and the music hold, against the five-dollar cap. A response with no usable usage tokens consumes at least the whole ceiling. A failed call is not retried. The refusal record keeps the HTTP status and the response text. With this route on, when the server has a fal key and the quote or the reserve does not fit, the plan fails and does not use the OpenAI planner. When the server has no fal key, the OpenAI planner remains the fallback. Since 2026-11-07, a server with that fal key and the flag on can plan a video with no attached media without `OPENAI_API_KEY`, and Studio shows planning as available. A still format or a plan with attached media still needs `OPENAI_API_KEY`. Without it, that plan returns `MOTION_NOT_CONFIGURED` before it reserves credits. Sampled frames and a scene revision can use `openrouter/router/vision` only after each frame is at most 1,080 by 1,920 pixels and that ceiling is reserved. When the caller also passes reference frames, the review reserves a second call of at most 20 images before the output call. If either ceiling does not fit, neither call is sent. A caller that supplies reference URLs and output URLs sends one vision request for each after both ceilings are reserved. More than 20 images is refused before that request. The Studio plan does not supply those URLs. A packet vision review runs after the image and music steps. Each round takes frame samples from the composition beats. It scores them only when the caller passes a vision score port inside the budget; the Studio plan passes none, so it records one unscored round. Each round keeps its scene snapshot; the plan uses the best round's scenes, audio and ducking. The first-cut gate can still revise those scenes before the plan is saved; audio and ducking stay from the best round. A failed revise step is charged as failed and halts the loop. Sol stays unused. Before a packet plan starts any image job, a free check reads the planned scenes: real product pictures, no on-screen words in an image prompt, and the skill demo template. A plan that fails one of these fails with the reasons and starts no image job. Before Studio shows a packet first cut, a free quality check scores it against seven items: real product pictures, motion beyond fades, a music bed, one narration take, no on-screen words in an image prompt, no slide-deck layout, and the skill demo template. A skill with no music bed does not need one. Generated footage film may use one picture for each scene in one layout. A still that is still generating counts as an image-model still. The check makes no paid call. It changes each fade or missing cut to a moving cut that the style allows, and scores the video again. A packet skill does not need a product screenshot, and an attached picture does not change that. When an item still fails, the plan fails with the reasons, and Studio does not show that first cut. A skill with no packet is unchanged.

Since 2026-11-01, video plans with attached media stay on the existing OpenAI planner, including when `MOTION_OPUS_PLANNER_ENABLED` is `true`. This preserves the asset IDs, images, clip metadata and transcripts. Such a plan with no configured `OPENAI_API_KEY` fails before a provider submission. Text-only plans keep the flag behavior described above.

The vision budget records at least the larger of the quote and the reported provider cost. A reported overrun stays in `polish.ledger.spentUsd` or `failedUsd`, even when it exceeds `capUsd`; it is evidence, not permission to spend again. Further paid requests stop when the remaining budget cannot cover them. A paid reply with an unusable score or output envelope still counts when its cost can be read. Cost arithmetic keeps finite overruns instead of replacing them with the estimate. The free music check counts the union of audible track intervals inside the video timeline. The check uses the renderer’s fades and ducking. Muted tracks, silent fade frames and overlapping intervals do not add coverage. A saved quality rejection keeps stable rubric causes, which Studio shows in English or Spanish. Older saved errors keep their existing fallback.

A scene can have the optional field `imageSubject`. A composition can have the optional field `polish`. `polish.pendingMusic` records a packet music bed job; `attached` is `true` after the editor adds the finished bed once. `polish.selfCheck` records the quality check: `verdict` (`pass`, `pending` or `blocked`), `revised` and `failed`, which list rubric item IDs. `styleSpec` can include the optional `packet`. Keep `imageSubject`, `polish` and `packet` when you send the whole composition to `edit_project`. A composition without them plans and renders as before.

AI hook and text proposals on a styled video get its headline length, headline case and tone. The animation writer gets the style name, the headline length, the tone, the text enter frames and the full palette. These are facts for the model. A new hook scene gets the motion and text position of the style. A skill with no typed style, and a video with no `styleSpec`, plan and render as before, with the same colors.

When styles are off, every skill plans with no style. A composition that adds a `styleSpec` or changes the saved one returns `INVALID_INPUT` from `create_project`, `edit_project`, a Studio save (`PATCH /api/motion`) and a Studio create (`POST /api/motion` with a composition) when styles are off, or when `VIDEO_SKILL_STYLES_ENABLED` is exactly `true` and the owner is not listed in `SHIPCONTENT_VIDEO_SKILLS_USERS`. The server reads the saved style. A save compares with the saved revision that it replaces. A convert, from `convert_project` or from the Studio convert dialog, compares with its source revision in the same workspace. A new project without a source has no saved style. A composition can keep or remove its saved `styleSpec`. A restore of an earlier revision keeps the style of that revision. A create sent again with the same request ID and content returns the saved project. A video saved with a style keeps it and renders with it.

### Skill run service

Updated 2026-11-05: The owner routes below record and manage runs for published recipes. They make no paid provider call and charge no credits. They do not enable a catalogue entry. The current availability is described under [Video skills](#video-skills). Run limits are reversible service defaults, not a price or an entitlement.

Every run request needs `?workspaceId=<uuid>`, a signed-in owner session and current access to that workspace. POST requests also need the same site origin and the session's `x-shipcontent-csrf` token. These routes reject agent authorization headers. There is no skill run MCP tool in this release.

| Request | Input and result |
| --- | --- |
| `POST /api/skill-runs` | Send `skillId`, a UUID `requestId` and `inputs`. Returns the run. |
| `GET /api/skill-runs` | Optional `filter`: `unfinished`, `ended` or `all`. Use `limit` and the last run's ID as `before` for the next page. Returns runs, newest first. |
| `GET /api/skill-runs/{id}` | Returns `{ run, links }`. Available file links are `preview`, `stills` and `video`. |
| `POST /api/skill-runs/{id}/answer` | Send the current `previewId` and `answer`: `approve` or `changes`. For changes, send the complete replacement `inputs`. Returns the run. |
| `POST /api/skill-runs/{id}/cancel` | Returns the cancelled run. Repeating cancellation returns that run. |

`inputs` contains `screenshots`, `headline`, `captions` and `lengthSeconds`. A screenshot names either a Library `assetId` or a ready `runInputId` from the same workspace. The service checks the real image bytes against the recipe's input limits. It saves the recipe, kit version and workspace brand when the run starts.

The `captions` array may have fewer entries than `screenshots`, or use `null` for a slot; missing and null slots use the per-screen default label at render time. It may not have more string or null entries than `screenshots`. Each non-null caption must stay within the recipe's per-caption character limit.

Keep the original `requestId`, `skillId` and `inputs` after a lost response. Replay them unchanged to read the saved run, including its later change rounds. This works after trial removal or recipe unpublishing, while workspace access remains valid. Reusing that request ID with a different skill or inputs that differ from the original request returns [`SKILL_RUN_CONFLICT`](/docs/contracts/errors#skill-runs). A new run needs a new request ID.

A new start requires a fresh, available renderer check-in. Once queued, only renderer downtime counts toward the claim-wait limit. Repeated unavailable check-ins and late usage reports do not restart that wait. Renderer recovery pauses it. Returning to the queue after an answered review starts a new wait. Reading or listing expires a queued run whose claim wait is exhausted. Reading or listing also closes a review run when its current stop expired unanswered. Follow `checkAgainAfterSeconds`; it uses the same downtime clock while queued and unavailable, and the review deadline while waiting for an answer. It is null for an ended run.

At review, read `stop.previewId`, `stop.expiresAt`, `round` and `maxRounds` before answering. The first recorded answer wins. An expired stop refuses both approval and changes with [`SKILL_RUN_STALE_ANSWER`](/docs/contracts/errors#skill-runs) without closing the run; the next read or list applies the expiry sweep and sets [`NO_ANSWER_IN_TIME`](/docs/contracts/errors#skill-runs). Approval queues finalization; it does not itself mean the video is saved. Read until `state` is `finished` and `libraryAssetId` is set. See [skill run errors](/docs/contracts/errors#skill-runs) for stale answers, capacity refusals and other failure codes.

File links at `/api/skill-run-files/{token}` stream private bytes without a redirect or a login cookie. Treat each link as a secret. Each read checks the named owner's membership or agent grant and account access again. Invalid, expired or revoked links return `NOT_FOUND`. Read the run again for fresh links. Run and file responses use `Cache-Control: private, no-store`.

## Video projects

Updated 2026-11-19: a video project also needs the owner in `SHIPCONTENT_STUDIO_VIDEO_USERS`.

Updated 2026-11-13. A video project is on only for the owners in the server setting `SHIPCONTENT_STUDIO_PROJECT_USERS`. For other owners, `?video=` on the Studio page and the routes below return `NOT_FOUND`. A project has four steps: idea, script, storyboard, and video.

1. **Idea.** Open Studio and select **Video project**, or go to `/{workspace}/studio?video=new`. Write what the video is about, or paste your script. Select images from the Library and what each image shows. Select the length (5 to 30 seconds), the size (9:16, 4:5 or 1:1), and the quality. **Write script** is a paid AI call.
2. **Script.** Edit the shots and the characters, places and products. Each edit saves a new version on its own and costs nothing: **Done** on a shot or cast member, **Apply changes** on the whole text, and the ending line and holds when you leave the field. There is no Save script button. The trash icon on a shot card asks you to confirm, then removes the shot and saves at once. The video keeps its length: the other shots share the removed time in proportion to their own lengths, in steps of half a second. **Undo** in the notice puts the shot back for 5 seconds. The last shot cannot be removed. **Rewrite shot** and **Describe again** are paid AI calls on one shot or one cast member.
3. **Storyboard.** **Make storyboard** makes one picture of each cast member in the shots, then one picture of each shot (or one sheet). The button shows the total price. A picture that failed returns its credits; make it again from its tile.
4. **Video.** **Make video** makes the video in parts, with the voiceover and the music when you select them. The parts, the voiceover and the music have one total price on the button. A take with all parts done becomes the project's video, and the project opens in the editor. A failed part is not sent again by itself: **Make again** starts it again for its own price.

The page calls owner routes. They need the signed-in browser session and a CSRF token. An agent token cannot call them. MCP tools do not show a project until its first take is done.

- `GET /api/motion/projects` lists the projects of the workspace.
- `POST /api/motion/projects` starts a project and writes its script. Send the same `clientRequestId` again after a lost reply.
- `GET /api/motion/projects/{id}` reads the project. The read also checks running jobs and sends waiting ones, at most 10 for each read.
- `POST /api/motion/projects/{id}` runs one action on the project.
- `GET /api/motion/projects/{id}/quote` returns the price of each next step.

A paid action sends the total that the owner saw. If the prices changed, the action returns `PRICE_CHANGED` and starts nothing. An action on an old script version returns `STORY_CHANGED`. Music needs the server music setting.

## The editor

The editor has three columns. The left column adds things: **Library** (images, clips and audio), **Scenes** (title, closing card, image, **UI walkthrough** for a video, and animations) and **AI**. The middle column is the preview with the scene or slide strip under it. The right column, the Inspector, edits the selected scene, slide or the whole project, with **Ask AI** at the bottom. The full timeline and animation tools need a screen of 1024 px or wider. Without first-video controls, a narrower screen shows preview and export only.

Updated 2026-11-14: in a video project, each script edit saves on its own. The Save script and Discard edits buttons are removed, and a failed save shows the last saved script again. A shot card has a trash icon that removes the shot after a confirmation, with a 5-second undo. A click on a character, place or product picture opens it large. On a character, place or product card, **Make picture** or **Redo picture** lets you pick the image model (GPT Image 2.5 Flare, GPT Image 2.5 Sunburst or GPT Image 1 Mini) with its price. **Use my image** picks or uploads an image from your library and uses it as the picture, at no charge. **Add reference image** gives the model an image to start from; it saves with the script and costs nothing.

Updated 2026-10-31: when Spanish is available for your account, the video editor shows your account language, English or Spanish, in its panels, buttons, notices, the Sizes view, Convert, Remake, the export view and the post step. Changing the language does not reload the editor. Your selection, undo history, playback and unsaved text stay as they are, and the change sends no save, render, generation or transcription request. Counts, times and credits use the number format of the language.

Text that you wrote or that the project holds stays as written: scene text, captions, layer text, the project title, saved project data and template names. A saved default such as the scene text of a new scene stays in English until you edit it.

A saved transcription failure that the app knows shows in your language, and one that it does not know shows a general sentence in Spanish. Reading a saved failure never sends a request again. In the skill list, the reference group (**Instructions only**) and the technique notes stay in English. The text that the editor sends to a provider or an agent does not change with the language.

With those controls enabled, **Edit video text** appears below the phone preview. It starts at the first ordinary scene and seeks to that scene's start. Select another scene to seek to its start. Headline or closing text allows 80 characters; supporting text allows 160. Video duration uses half-second steps from 0.5 to 30 seconds per scene, within 30 seconds total and the source clip's available length. Other scenes and words stay unchanged. The same editor save and undo path is used. **Project name** remains separate. A generated animation directs you to a larger screen; its basic text fields are not editable on a phone. Resizing keeps the same preview and edits.

The top bar has the title, the save state, **Undo**, **Redo**, **Versions**, the project settings, the size switch and **Sizes**. The editor saves by itself. It also saves when you leave the editor, unless a save failed or a conflict is open. **Versions** lists saved versions; a save by an agent shows "agent". Restoring a version saves it again as the newest version. If an agent or another window saved first, the editor shows **Discard my edits and load the latest version** and does not overwrite that save.

With first-video controls, **Return to post** pauses editing, waits for pending native operations and confirms that the last accepted edit is saved before leaving. An unchanged project with a confirmed saved version needs no new save. Save errors, conflicts or an unavailable return target keep the editor open. If a save or restore may have completed without confirmation, undoing to the old saved text does not resolve it. Use **Retry** for a save error, or **Discard my edits and load the latest version** for a conflict; loading replaces the local draft. A delayed restore or load cannot replace edits accepted while it was pending. **Cancel return** or **Stay** releases editing and retains any unresolved manual-create attempt. The [manual-create exit warning](#start-a-project) still applies while Return waits for an answer.

- Select a scene in the strip to change its length, start in the clip, text, motion and clip sound. Since 2026-11-07 a length that is not on the half-second grid, for example a length that the narration fit set, shows as a valid value, also at the maximum of the field, and the half-second steps count from it. Since 2026-10-31 **Wipe** and **Push** appear only for a generated scene in a packet video, or for a scene that already uses one. Move, duplicate or delete it there.
- Use the size switch to preview each size. **Adjust for this placement only** changes the frame, focus and text position for one size.
- Add walkthrough layers to an image or clip in a video: **Highlight**, **Callout**, **Cursor** and **Zoom**. A scene holds up to 8 layers. On a still slide, **Highlight**, **Callout** and **Zoom** stay fully visible. Adjust or remove them in **Walkthrough layers**; still slides have no layer timing controls.
- For an animation, change its fields. **Edit with chat** also changes it with an instruction.

### AI proposals

**Ask AI** rewrites the words of one scene or slide (**This scene**) or of the whole project (**Whole project**). Write an instruction of up to 500 characters and select **Send**. The editor does not change the project. The editor shows the result as a proposal. The preview shows the new text. **Proposed changes** lists the old and new words. A bar under the preview shows **Previewing proposal** with **Discard** and **Apply**. Nothing is saved until **Apply**. The proposal stays open while you edit the project, and the preview applies it again on top of your edits. If an edit or an **Undo** leaves the proposal nothing to change, the proposal closes and your edit stays. **Apply** is one undo step, and **Undo** reverses an applied proposal. **Undo** does not close an open proposal: it reverses your last edit and keeps the proposal. **Redo** is not available while a proposal is open. **Discard** closes the proposal. The request goes to the owner route `POST /api/motion/text` and sends the project's text to OpenAI, a paid model call. The request is saved before the call: a repeat returns the saved result, and a workspace runs at most two of these calls at a time. The server checks every scene ID. If the source revision changes while the provider works, the completed request closes without saving a proposal. Its status states whether credits were charged. Starting again creates a new request. Without the server OpenAI key, the route returns `MOTION_NOT_CONFIGURED`.

On a still slide, **Write text from facts** makes the same kind of proposal. For an animation scene, **Ask AI** uses the animation writer, and its result is a proposal too.

### Sizes

Select **Sizes** to see the selected scene or slide at every size that the format allows, with its pixel size and the channels that accept it. Video scenes show their middle frame. Still slides and generated animations show their completed reveal. Select a card to edit the layout of that size. A card marked "edited" has its own layout. Each size card warns when the selected scene's text overflows that size or enters the safe area. The measure is an estimate from the text length and the layout tokens, and it includes the chrome the preview already draws. It is not a live fit check. **Fix: smaller text** steps only that size from 100 percent to 90, then 80, then 70 percent. At 70 percent the button stops, and a warning can remain. Animation text and walkthrough layers are not measured.

### Convert

A project keeps its format. **Convert to Carousel**, **Convert to Story** and **Convert to Video** in the project settings make a new project from the current version. The source project does not change. Convert is free and calls no AI.

- A video with 2 or more scenes can become a carousel. The first 10 scenes become slides. Highlight, Callout and Zoom become static; Cursor is removed. A video clip becomes a slide with its middle frame; without a frame, it becomes a text slide. The dialog lists what it did not include.
- A video can become a Story. The Story keeps the scenes.
- A carousel can become a video. A conversion to a carousel or to a video removes audio and captions.
- A single image and a Story do not convert.

### Add audio and captions

- In **Audio**, add music, a voiceover or a sound from the Library, with volume and fades. A video holds up to 6 tracks. **Lower music while the voiceover plays** reduces the music under a voiceover.
- Under **Voiceover**, write a script of up to 2,000 characters, select a voice and select **Create**. The action makes an audio file in the Library. This is paid; the panel shows the credits. Select **Add** to put the file in the video. Updated 2026-11-11: the voice list includes the additional stock voices named in the limits below. When Settings has a connected ElevenLabs account, the picker also lists voices from that account and narration with one of them uses the owner key. A refused owner voice fails with `MEDIA_PROVIDER_REJECTED`. The skill start form still offers only built-in voices. See [ElevenLabs owner account](/docs/contracts/errors#elevenlabs-owner-account).
- In the video settings, **Captions** sets the style and position. For a clip with speech or a voiceover, select **Transcribe**. Transcription returns the words. This is paid. Its quote quantity is the saved duration rounded up to a whole second, capped at 600 seconds; the catalogue applies its minute block to that quantity. Updated 2026-10-29: the server sends only these quoted seconds of sound to the transcriber. Sound after them gets no words. A missing or invalid saved duration stops the request. A ready or running transcript for the same asset is reused without another paid submission. An unknown submission is not sent again. If an approval expired before provider submission, an explicit check closes that old request; another owner approval can create a new request. A failed request can be replaced only by a new owner action. Correct a word in the list. The export burns the captions into the video. Since 2026-11-07, in a packet video the spoken captions take the place of the fixed caption line of each scene. That line keeps its space but does not show. On a generated scene with a ground too light for white words, **Clean** captions use the text color of the scene and no dark shadow. When the voice starts within the first second, the first caption line shows from frame 0, so the cover frame has words.
- On a transcribed clip, select **Remove pauses and fillers**. Review the proposal, then select **Apply**. The clip becomes one scene for each part it keeps. **Undo** restores the clip.

### Remake with AI footage

1. Select a scene and select **Remake with AI footage**. **Send to Seedance** and **Use this frame** open the same drawer for a clip range or one frame. **Add b-roll after** is the same drawer for a shot that starts from a prompt only.
2. Set the range (2 to 15 seconds), the picture, the sound in the new clip, the prompt, the model, the quality and the length. Keep **Clean picture** on. It leaves out the text, the logo, the captions and the highlight, callout and cursor layers, because Seedance can distort words and logos.
3. The drawer shows the credits. Select **Generate**. The worker renders the reference clip or frame first. This render is free. Then the paid generation starts.
4. When the shot is complete, it shows under **Seedance results**. Select **Replace scene** to put it in the origin scene, or **Insert after** to add it after that scene.

After **Replace scene**, the scene keeps its text and layers on top of the new shot. Text inside the Seedance footage has no protection.

### Brand kit and templates

The brand kit of a workspace has a name, a main and a second color, a logo image and a caption style. Without a saved kit, the page suggests one from the latest website import; select **Save as brand kit** to keep it. In a project, **Use brand kit** applies the kit, and **Save as brand kit** saves the project's brand as the kit. A logo must be an image.

**Save as template** keeps the scenes, lengths, layouts, text and caption settings of a project, but no media. A workspace keeps up to 30 saved templates. The built-in templates are Product launch, UGC testimonial, Product UI walkthrough, AI footage ad, Kinetic typography, Data story and Technical explainer.

## Export

Select **Export & publish**. The view lists every size the format allows. With first-video controls enabled, it carries the current editor size and selects only that size first; an absent or invalid size uses the first allowed size. Other sizes require explicit selection. **Create MP4** requests one video size; **Create MP4s · N sizes** names a batch. Images keep their format-specific export labels. Local selection stays while this export view is open and resets for a different project. Since 2026-11-07, the sizes rendered for the open revision stay selected when export opens again, for example 9:16 only; an open revision with no render keeps the first selection. Since 2026-11-08, when the first render of the open revision arrives after export opens, its sizes become selected, unless you changed the selection. The server renders the saved revision; each requested size is its own job. Existing files remain available when a size is unchecked.

- A video renders one MP4 at 30 fps for each size.
- A carousel, a single image and a Story image render one JPEG for each slide for each size.

Each size has its own state: **Not rendered**, **Waiting for approval**, **Queued**, **Rendering**, **Ready**, **Failed** or **Status unknown**. A failed size shows **Retry** and **Skip**. Read the current credits per render before starting work; if the quote cannot be read, the view shows **Render price unavailable**. With first-video controls, **Status unknown** offers **Check status**, which does not start a replacement render. **Checks before render** shows warnings. Examples: a video that is too long for a channel, text in a Story near the app controls, or a carousel with a slide count that a channel does not accept. The checks also warn when scene text overflows a size or enters a safe area, and when a real caption line does. Empty captions do not fail. The measure is an estimate. At 70 percent the text can still warn. Animation text and walkthrough layers are not measured. Since 2026-11-07 a text check that measured nothing shows a dashed circle, not a check mark. The checks do not block the render. Each file goes to the Library.

Updated 2026-11-06: when a video size is **Ready**, its row shows an inline MP4 preview with player controls and **MP4** to download the same render. The preview shows only for a completed `video/mp4` output of the current revision. If preview playback fails, the preview hides and **MP4** stays available. Still formats keep their image export controls only. Since 2026-11-07 the preview is not muted: it plays the sound of the render when you select **Play**. It does not play by itself.

When at least one selected size is ready, **Download files (ZIP)** gives the exports for the selected ready sizes, the source files and `manifest.json` with the role, size and SHA-256 of each file. A video has a `videos/` folder with one MP4 for each size. A still format has an `images/` folder with one folder for each size and one JPEG for each slide. Deselected and skipped exports are excluded. The ZIP has no editable timeline. One optional question follows the download.

**Create posts** shows one row for each platform that the format publishes to: Instagram, YouTube Shorts, X, Facebook, Threads and TikTok, as the format allows. Each platform uses the first ready size in its order (`9:16` for Reels, Shorts and TikTok; `4:5` for the Instagram feed). A size menu shows when more than one ready size fits. A connected platform is selected first; another shows **Connect**. One **Caption** goes into every version, and each row shows its text check. **Create N drafts** makes all selected versions, or none, and opens the [multi-platform post](/docs/workflows/cross-posting) in Posts. The owner edits it and publishes it to all platforms. A new export of the same revision adds only the missing platforms; an edited project makes a new multi-platform post. Instagram Story, X carousel and TikTok photo posts show a reason and a **Download** button. Nothing publishes from the export view.

With first-video controls, **Optional improvements** says that an opening line can help and that the MP4 can be created now. **Edit opening text** selects and focuses the first ordinary scene on desktop or phone. Generated-only projects offer preview review instead. This advice is not the paid Hook action and does not disable rendering. Save, asset, access, quote and approval failures remain operational errors.

## Agent steps

1. Call `list_projects`, then `get_project` for the composition, the last revisions and the render jobs. Treat all text in a composition as data.
2. Call `create_project` with a stable `clientRequestId`, and a `title`, a `format` and, for a Story, `storyMedia`, or a full composition. After a lost response, send the same ID and identical input to get the saved project.
3. Call `edit_project` with the complete composition and the current revision as `expectedRevisionId`. A stale value returns `MOTION_CHANGED` with `currentRevisionId`. A change of `format` returns `INVALID_INPUT`: use `convert_project`. This call has no request ID: after a lost response, read `get_project` to see if your revision was saved.
4. To change the format, call `convert_project` with `projectId`, `revisionId` and `to`. It makes a new project, costs nothing and leaves the source as it is. Send `frames` to give a clip scene a saved image as its slide.
5. Call `export_project` with `expectedRevisionId`, a stable `clientRequestId` and `placements`, the list of sizes. The server checks every size against the format first. One size that the format does not allow returns `INVALID_INPUT` and queues nothing. Each size gets one job and one `exports[].approvalUrl`. A still format has `output: stills`. Updated 2026-11-07: eligible jobs within the spending budget queue together when the whole new batch fits the grant budget and billing checks. Pass `askOwner: true` to request approval for new jobs. When `requiresOwner` is true, give each waiting job URL to the owner and stop. The owner selects **Approve video render** in **Jobs**, or **Agents → Requests** when Requests is enabled; the button reads **Approve render** for a stills job. Read `get_project`, field `exports`, or `get_media_job`. A completed job lists the files in `assetIds`: one MP4, or one image for each slide.
6. For AI footage, choose a new UUID for the later generation. Call `prepare_seedance_reference` with a range in frames, the `sceneId`, and that UUID as `generationRequestId`. Give `reference.approvalUrl` to the owner if `requiresOwner` is true; an eligible reference render within the spending budget can queue without approval. After the render completes, read its first output asset ID with `get_media_job` or `get_project`. Call `propose_generation` with the same UUID as `input.clientRequestId` and that asset ID in `input.referenceIds`: reference-to-video for a clip, image-to-video for a still. That paid generation has its own spending budget or owner approval decision. Check its `requiresOwner`. The owner then sees **Replace scene** and **Insert after** for the result. You can also put the new asset in a scene with `edit_project`.
7. For a voiceover, call `propose_generation` with mode `text-to-speech`, model `elevenlabs-tts`, the script in `prompt` and an optional `voice`. After its owner-approved or spending budget job completes, add the audio asset as a `voiceover` track with `edit_project`.
8. To publish, make a post draft with `create_post` or `edit_post` and attach a completed export, then continue with the channel flow. No project tool publishes.

Agent render jobs wait in `awaiting_approval` unless eligible spending budget queues them. All unfinished jobs count toward the ten-job workspace limit. The spending budget authorizes new work; an agent cannot approve an existing waiting job. Replays preserve saved jobs and add no holds. See [media decision fields](/docs/workflows/media#propose-generated-media) for `reason`, `budgetLeft` and `creditsCharged`. A signed-out owner who opens a project link signs in and returns to the same project or render job. An owner link from a tool is `/studio?workspaceId=<id>&project=<id>`, with `mediaJobId=<id>` for a render job.

## Limits

- A video has 1 to 12 scenes, 15 frames to 30 seconds each, and 30 seconds in total. A carousel has 2 to 10 slides.
- Up to 6 audio tracks, and up to 8 walkthrough layers in a scene.
- A Seedance reference clip is 2 to 15 seconds. Seedance accepts video references of 2 to 15 seconds and less than 50 MB in total.
- A voiceover script has up to 2,000 characters. Updated 2026-11-02: the script can hold Eleven v4 audio tags in square brackets, such as `[excited]` or `[short pause]`. Each tag counts as characters. The AI first cut can add a few tags to its script. The voices are Jessa (default), Clara, Ana María, Alec, Corey and Mario.
- Updated 2026-11-02: the old voice names Sarah, Jessica, Alice, Brian, Eric and Charlie are still accepted. They map to Jessa, Jessa, Clara, Corey, Alec and Mario.
- Updated 2026-11-10: the voice list also includes Elara, Talia, Florence, Maisie, Jade, Alicia, Eldrin, Finley, Lawrence, Wyatt, Caleb and Baxter. Jessa stays the default. The old names still map as they did on 2026-11-02. These are existing ElevenLabs stock voices. The product does not create, clone or design a voice.
- Updated 2026-11-12: a connected owner account shows at most 500 voices. A voice name longer than 100 characters is cut. ShipContent keeps the voice ID and that name only.
- A first cut uses a brief of 10 to 4,000 characters and up to 6 images or clips. With a [video skill](#video-skills), the brief is optional and has up to 4,000 characters.
- Updated 2026-10-29: planning reads 3 frames of each first-cut clip. If a frame that it reads has more than 16,777,216 pixels (4096×4096), planning refuses the clip with `MEDIA_LIMIT_EXCEEDED` before the paid model call. Choose a smaller clip.
- An **Ask AI** instruction has up to 500 characters and covers up to 12 scenes.
- Up to 30 saved templates in a workspace.
