| name | gemini-gen |
| description | Structured image and video understanding via the native Renoise CLI and Gemini 3.1 Pro. Extract standalone generation prompts, reusable replacement templates, timelines, dialogue, camera language, style, subjects, audio, and uncertainty warnings. Use for understanding local images/videos, reverse-prompting, script/style extraction, and preparing media for Director workflows. Do NOT generate media here — use renoise-gen.
|
| allowed-tools | Bash, Read |
| metadata | {"author":"renoise","version":"0.4.0","category":"ai-foundation","tags":["vision","analysis","multimodal","gemini"]} |
Analyze Media — Native Renoise CLI
renoise analyze is the only runtime for media understanding. It uses Gemini 3.1 Pro, validates structured output, streams large files through a temporary analysis upload, and never saves the source as a generation material or creates a paid media-generation task.
Preflight
Verify the native command and authentication:
command -v renoise
renoise help analyze
renoise auth status --json
On Windows PowerShell use Get-Command renoise. If any check fails, immediately read ${CLAUDE_SKILL_DIR}/../renoise-setup/SKILL.md completely and follow it through readiness, asking only for approvals required before host changes or browser authorization. Then rerun preflight and continue the original request; do not merely direct the user to Setup / Account.
Core Commands
renoise analyze photo.jpg --target image --language <user-language> --json
renoise analyze photo.jpg --target video --language <user-language> --json
renoise analyze clip.mp4 --target video --language <user-language> --json
renoise analyze reference.mp4 --mode template --target video --language <user-language> --json
renoise analyze clip.mp4 --prompt-only
The default mode is standalone; its prompt contains no placeholders or reference requirements. template preserves composition/timing/style while replacing source-specific identities with model-neutral {{slot_id}} placeholders. Each slot includes a self-contained English reference-image prompt.
Result Contract
Use the structured fields rather than reparsing the generated prompt:
{
"version": "v1",
"model": "gemini-3.1-pro",
"mode": "standalone",
"target": "video",
"source": {
"name": "clip.mp4",
"type": "video",
"mimeType": "video/mp4",
"size": 123456
},
"analysis": {
"summary": "...",
"durationSeconds": 8,
"aspectRatio": "9:16",
"composition": "...",
"facePresence": "present|absent|uncertain",
"timeline": [
{
"start": 0,
"end": 2,
"visual": "...",
"action": "...",
"camera": "...",
"dialogue": "...",
"sound": "..."
}
],
"style": {
"lighting": "...",
"palette": ["..."],
"cameraLanguage": "...",
"pacing": "...",
"mood": "..."
},
"subjects": [{"id": "subject_1", "type": "character", "description": "..."}],
"audio": {"dialogue": "...", "music": "...", "effects": "..."}
},
"prompt": "complete English generation prompt",
"slots": [],
"warnings": []
}
Schema keys and the final generation prompt are English. Analysis descriptions follow --language; dialogue remains verbatim in the detected source language.
Common Tasks
Script and Shot Extraction
Use analysis.timeline for timestamped scenes, actions, camera movements, dialogue, and sound. Use analysis.audio for aggregate dialogue/music/effects. Timestamps are model observations, not frame-accurate edit decision lists; preserve warnings and verify critical dialogue with the user.
Style Extraction
Use analysis.style plus analysis.composition. Do not infer style by scraping adjectives from prompt when structured fields are available.
Product or Character Understanding
Use analysis.subjects, analysis.summary, analysis.composition, and analysis.facePresence. uncertain is not equivalent to absent; any privacy-sensitive automation must fail closed.
Compare Two Assets
Analyze each local asset separately with the same target and language, then compare their structured analysis objects. Do not upload either as a generation material unless the downstream workflow explicitly requires it and the user approves.
Reference-Video Remake / 剪同款
Route to the director skill and read commercial/scenario-a-viral.md. That workflow uses --mode template, persists remake-analysis.json and remake-plan.json, fills replacement slots, checks live model role combinations, and requires approval before slot generation and final video generation.
Input and Failure Rules
- Supported local files: jpg/jpeg, png, webp, gif, mp4, mov, webm.
- MIME/content mismatches are rejected.
- Large media uses a temporary analysis URL; it is not added to the Renoise material library.
- Cross-media output (
image → video, video → image) records inferred details in warnings.
- If output is truncated, malformed, blocked, or missing required fields, treat the command as failed; never fabricate a replacement result.
analyze may consume LLM analysis quota, but it never creates a media generation task.
- For generation, pass the approved
prompt into renoise task create through the renoise-gen or director workflow.