ai-multimodal_skill

Overview

This skill provides a practical toolkit to analyze and generate images, audio, and video using Google Gemini multimodal models. It combines preflight media optimization, automatic API key rotation, and a CLI orchestrator to run analyze, transcribe, extract, and generate workflows. The goal is fast, reliable multimodal AI for vision analysis, transcription, OCR, image/video creation, and document conversion.

How this skill works

The skill inspects input media (images, audio, video, PDFs) and routes them through sensible Gemini models and presets. It offers a CLI orchestrator that auto-selects models, streams files inline or via the File API, and saves structured outputs (text, JSON, CSV, Markdown, generated assets). Complementary tools preprocess media (ffmpeg/Pillow) to meet Gemini limits and handle key rotation under rate limits.

When to use it

• You need high-quality image analysis, OCR, or visual question answering on screenshots, product photos, or documents.
• You want transcriptions or speaker-segmented transcripts for meetings, podcasts, or interviews.
• You need to generate images with Imagen 4 or short videos with Veo 3 for concept art or marketing.
• You must batch-process large media directories while respecting Gemini size and duration limits.
• You want an automated pipeline that handles file conversion, chunking, and reliable API key rotation.

Best practices

• Preflight media with the optimizer to resize/compress and avoid inline size limits (20 MB) when possible.
• Split audio/video into <=15 minute chunks for full transcripts and to prevent response truncation.
• Enable multiple API keys for rotation if you run high-volume jobs to reduce rate-limit failures.
• Choose task-appropriate models: gemini-2.5-flash for fast analysis, imagen-4 for image gen, veo-3 for short video generation.
• Save structured outputs (Markdown/JSON/CSV) to make downstream automation and auditing easier.

Example use cases

• Extract structured data and OCR from multi-page PDFs into markdown-preserving documents for documentation teams.
• Batch transcribe and timestamp long podcasts by splitting audio, transcribing segments, and merging transcripts.
• Analyze product screenshots to produce captions, UI element lists, and accessibility suggestions.
• Generate concept images for campaigns using imagen-4 with quality/speed presets, then iterate with prompts.
• Produce short 8s preview videos with Veo for social clips or animated storyboards.