Remotion Production Workflow

This skill covers what Remotion's own skills don't: getting the media. It orchestrates 5 MCP servers to produce voiceovers, music, SFX, footage, images and clips, then wires them into a Remotion composition.

Works on top of the official skills

This plugin builds on each provider's own Agent Skills and keeps only what's specific to making Remotion videos:

Provider	Official skills	Install
Remotion	remotion-best-practices + sub-skills	npx skills add remotion-dev/skills
ElevenLabs	text-to-speech, sound-effects, music, speech-to-text, voice-isolator, voice-changer	npx skills add elevenlabs/skills
Replicate	find-models, run-models, compare-models, prompt-images, prompt-videos	npx skills add replicate/skills

TwelveLabs has no Agent Skills; its agent docs are at https://docs.twelvelabs.io (and llms.txt).

Remotion maintains the official Agent Skills (remotion-best-practices and its sub-skills: remotion-create, remotion-markup, remotion-studio, remotion-render, remotion-captions, remotion-multimedia, …). They're the source of truth for Remotion APIs and change with every Remotion release. If they're missing, install them first: npx skills add remotion-dev/skills (or run the setup skill).

When both are loaded, split the work like this:

Concern	Follow
Voiceover, sound effects, music	Official ElevenLabs skills: a small SDK script in the project that writes to public/ with this skill's names. Defaults: eleven_v4 (speech), music_v2_5 (music).
Images and AI clips	This skill: KIE via remotion-media; Replicate (with the official Replicate skills) for specific models or when KIE fails
Stock footage, footage analysis, AI review	This skill: Pexels and TwelveLabs MCP tools
Remotion markup: components, imports, timing props, animation, effects, transitions, fonts	Official remotion-markup
Project scaffolding	Official remotion-create
Preview	Official remotion-best-practices / remotion-studio: start Studio as soon as the project runs, before building scenes, and keep it open
Rendering	Official remotion-render: only when the user asks for an MP4, or when running review-video
Captions: word-level transcription and display	Official remotion-captions (this skill only adds TikTok-style presets and short-form placement)
Measuring media durations	Official remotion-multimedia (Mediabunny), inside calculateMetadata
Anything not covered here	Check the live docs via the official remotion-docs skill rather than guessing

If a code snippet here ever disagrees with the official skills, the official skills win.

Providers and MCP tools

ElevenLabs (voice, SFX, music: the default)

• Generate with the official ElevenLabs skills (SDK script, ELEVENLABS_API_KEY). Eleven Music is cleared for commercial use on paid plans.
• Deprecated alternative: the ElevenLabs MCP server (elevenlabs-mcp 0.12.2, archived upstream) still ships for one more release for quick interactive work: text_to_speech, text_to_sound_effects, compose_music, voice_clone, search_voices, search_voice_library, isolate_audio, speech_to_text. It only reads and writes inside public/; pass output_directory as the project's public/.

remotion-media (via KIE: images and clips)

• generate_image: AI images (Nano Banana Pro)
• generate_video_from_text: AI clips from a prompt (Veo 3.1)
• generate_video_from_image: animate 1 image, or transition between 2. image_urls must be public URLs.
• generate_speech, generate_music, generate_sound_effect: drafts only. Suno and the other KIE-resold models aren't licensed by their owners through KIE. Use KIE audio for drafts and internal work, and generate final commercial audio with ElevenLabs.
• list_assets: browse the Airtable asset library (requires AIRTABLE_API_KEY)
• Don't use its subtitles tool; captions go through the official Remotion flow.
• KIE's Veo depends on a preview Google is retiring. If a clip call fails, fall back to Replicate.

These tools return the file path, not the duration. Measure durations as the official remotion-multimedia skill shows.

TwelveLabs (video understanding and AI review)

• create-asset (upload a local file) → sync-analyse-video or async-analyse-video with modelName: "pegasus1.5" and a JSON response schema → delete-asset when done
• search only finds moments inside an indexed library ("the part where…"); it doesn't analyse.

Pexels (stock footage)

• pexels_search_videos, pexels_search_photos, pexels_popular_videos, pexels_curated_photos, pexels_get_video, pexels_get_photo, collections
• No download tool: take the file URL from pexels_get_video (or the search result) and curl it into public/footage/.
• Pexels doesn't require attribution, but it's appreciated.

Replicate (optional, specific models)

• Local replicate-mcp with REPLICATE_API_TOKEN: search_models, get_models, list_models_examples, create_models_predictions / create_predictions, get_predictions, download_files
• Follow the official Replicate skills: search for the model and read its input schema before calling. See rules/replicate-models.md for a dated shortlist.

Asset conventions

• remotion-media tools write to public/<output_name>.<ext> in the directory the agent was launched from, so the agent must be started from the Remotion project root. They can't write into subfolders, so encode the type in the name: vo-scene1, music-bg, sfx-whoosh, img-hero, clip-city-aerial.
• Files you download yourself (Pexels, Replicate outputs) go into subfolders: mkdir -p public/footage public/images, then curl -L -o public/footage/<name>.mp4 <url>.
• Reference every file with staticFile() and its exact path relative to public/, e.g. staticFile("vo-scene1.mp3") or staticFile("footage/city.mp4").

Rules

Read the rule that matches the task:

• rules/production-pipeline.md: end-to-end order of MCP steps, from concept to preview
• rules/audio-integration.md: layering voiceover, music and SFX; ducking
• rules/voiceover-sync.md: per-scene voiceover with generate_speech, timed with calculateMetadata
• rules/music-scoring.md: music prompts, volume levels, per-scene music
• rules/sound-effects.md: SFX prompt library and timing to visual events
• rules/elevenlabs-advanced.md: voice cloning, voice library, audio isolation, multi-voice scripts
• rules/stock-footage-workflow.md: Pexels search, download, attribution
• rules/video-analysis.md: TwelveLabs indexing, search and clip selection
• rules/image-generation.md: image prompts and provider choice (KIE, Replicate, Pexels)
• rules/video-generation.md: clip prompts, provider choice, image-to-video pipeline
• rules/replicate-models.md: Replicate model catalog and when to use it over KIE
• rules/captions-workflow.md: TikTok-style caption presets on top of the official captions flow
• rules/animation-presets.md: text-animation recipes in the official markup style
• rules/data-visualization.md: animated charts, dashboards and counters
• rules/visual-effects.md: what to use from the official effects, plus letterbox and glitch
• rules/ci-rendering.md: GitHub Actions workflows for automated renders
• rules/asset-management.md: naming, public/ layout, staticFile() mistakes

For 3D, transitions, light leaks, Lottie, fonts and the other Remotion APIs, use the official remotion-markup skill directly.

Key principles

1. Audio drives timing. Generate the voiceover per scene first, measure each file in calculateMetadata, and size the scenes and composition from that. Durations in frames are integers: use Math.ceil.
2. Preview first. Start Studio as soon as the project runs, so the user watches assets land and can steer. Render only when asked, or for review-video.
3. Generate with tools, write markup with the official skills. MCP tools make the files; the official skills say how to place them (<Audio> and <Video> from @remotion/media, timing props, effects).
4. Preserve user changes. If the user edited the code between turns, keep their edits.