AI Generation

End-to-end deployment, evaluation, and management of AI agents on Microsoft Foundry. Covers the complete agent lifecycle: creation from starter samples, containerization and ACR push, hosted or prompt agent deployment, invocation, batch evaluation, and prompt optimization Includes specialized sub-skills for deploy, invoke, observe (evaluation and prompt optimization), trace analysis, troubleshooting, and dataset curation from production traces Supports project and resource provisioning, RBAC management, quota tracking, and model deployment with intelligent routing across regions and SKUs Requires .foundry/agent-metadata.yaml as the source of truth for environment-specific configuration, datasets, and evaluation test cases

Unified access to Azure AI services: Search, Speech, OpenAI, and Document Intelligence. AI Search supports full-text, vector, hybrid, and semantic search with AI enrichment capabilities like entity extraction and OCR Speech service enables speech-to-text transcription (real-time and batch), text-to-speech with neural voices, speaker diarization, and custom models MCP server integration provides direct tool access via azure__search and azure__speech commands; falls back to CLI and SDK when MCP is unavailable Includes OpenAI model access, DALL-E image generation, embeddings, and Document Intelligence for form extraction and OCR

Configure Azure API Management as an AI Gateway for models, MCP tools, and agents with built-in governance policies. Supports semantic caching (60-80% cost savings), token rate limiting, content safety filtering, and jailbreak detection across AI backends Add Azure OpenAI, AI Foundry models, or convert existing APIs to MCP tools as managed backends with load balancing Includes five core policy categories: authentication, semantic cache lookup, token limits, content safety, and token metrics for observability Requires Azure CLI for configuration and testing; integrates with managed identity for secure backend access

Build and deploy GitHub Copilot SDK applications to Azure with flexible model configuration. Three scaffolding paths: create new greenfield projects, add SDK services to existing repos, or deploy existing SDK apps with Azure infrastructure Supports three model configurations: GitHub's default models, specific GitHub models via discovery, or bring-your-own-model (BYOM) on Azure with DefaultAzureCredential authentication Includes complete templates with Express/TypeScript API, React/Vite frontend, Bicep infrastructure, Docker support, and token management scripts Deploy workflow uses azure-prepare, azure-validate, and azure-deploy steps; requires Docker and respects existing AGENTS.md configuration in user repos

Generate AI images with GPT-Image-2, FLUX, Gemini, Grok, Seedream, Reve and 50+ models via inference.sh CLI. Models: GPT-Image-2, FLUX Dev LoRA, FLUX.2 Klein…

Intent-routed video editing across three specialized RunComfy models. Routes user requests to Wan 2.7 Edit-Video (general restyle, background/packaging swap with motion preservation), Kling 2.6 Pro Motion Control (precise motion transfer from reference video), or Lucy Edit Restyle (lightweight identity-stable outfit/lighting changes) Each route includes model-specific schema, invocation examples, and prompting patterns to avoid iteration waste Requires RunComfy CLI, account login, and source video URLs (2–10s for Wan, 10–30s for Kling, no length cap for Lucy) Invokes models via runcomfy run <vendor>/<model>/<endpoint> and downloads results to a local directory

Smart router matching user intent to the right image-to-video model on RunComfy. Routes to HappyHorse 1.0 I2V (Arena #1) for general portrait and product animation with native audio synthesis Routes to Wan 2.7 with custom audio_url for lip-synced voiceover animation in multiple languages Routes to Seedance 2.0 Pro for multi-modal composition combining subject image, reference video, and reference audio Bundles model-specific prompting patterns and schema to optimize output without trial-and-error iteration Invokes via local RunComfy CLI; requires runcomfy login and a source image URL (JPEG/PNG/WebP, 300px min, ≤10MB)

Flash-tier text-to-image model optimized for rapid iteration, social thumbnails, and predictable in-image typography. Supports 11 aspect ratios (including 9:16, 21:9, 4:5) and four resolution tiers (0.5K to 4K) with tiered pricing; default 1K for cost efficiency Excels at subject-first declarative prompts with quoted text for reliable typography rendering and web-grounded image generation via optional enable_web_search Generates up to 4 images per request with seed locking for reproducible refinement across variants Routes to Nano Banana Pro for portraits, Flux 2 for heavy stylization, GPT Image 2 for maximum prompt adherence, or Seedream 5 for 2K–4K hero shots based on user intent

Precise single-image editing with Flux 1 Kontext Pro on RunComfy, optimized for high-fidelity local edits. Accepts a single source image URL and declarative edit prompt (e.g., "Keep the person's face unchanged, add an orange umbrella"), with optional seed for reproducibility Designed for targeted, preservation-focused edits; routes to Nano Banana Edit for batch workflows, GPT Image 2 for embedded text, or Flux 2 Klein for generation from scratch Requires RunComfy CLI ( npm i -g @runcomfy/cli ) and authenticated account; invokes runcomfy run blackforestlabs/flux-1-kontext/pro/edit locally Strongest results with single-instruction prompts led by preservation statements; compound edits should be split into sequential passes to avoid drift

Text-to-video generation with audio-driven lip-sync and multi-reference motion control. Supports audio-driven lip-sync via audio_url (WAV/MP3, 3–30s) for voiceover-synced videos, plus multi-reference conditioning for fine motion control across up to 5 reference media Generates 2–15 second clips at 720p or 1080p in five aspect ratios (16:9, 9:16, 1:1, 4:3, 3:4) with smooth motion physics and accurate transitions Includes prompt expansion by default to auto-rewrite short prompts; disable with enable_prompt_expansion: false for literal control Invoked via runcomfy run wan-ai/wan-2-7/text-to-video CLI; requires RunComfy account and npm i -g @runcomfy/cli

Image editing with OpenAI GPT Image 2, optimized for text rewriting, identity preservation, and multi-reference composition. Excels at preserving identity through targeted edits, rewriting embedded text in any script (Latin, kana, CJK, Cyrillic, Arabic), and layout-precise repositioning using directional language Accepts up to 10 reference images; first is primary, rest provide auxiliary cues for composition and styling Includes documented prompting patterns (lead with preservation goals, quote multilingual text, use numbered image refs) that produce sharper output than naive prompting Runs via runcomfy run openai/gpt-image-2/edit CLI; requires RunComfy account and token authentication

Fast image generation with Flux 2 Klein, optimized for sub-second iteration and multi-reference brand styling. Two variants: 4B for real-time concepting (sub-second at 4 steps), 9B for polished output (~25 steps); same prompt grammar across both Subject-first declarative prompting pattern yields sharper results than naive prompts; step-count strategy (4–8 for concepting, ~25 for polish) documented with anti-patterns Supports up to 4 simultaneous reference images for style transfer and guided composition without separate endpoints Runs via runcomfy run blackforestlabs/flux-2-klein/<variant>/text-to-image CLI; requires RunComfy account and token; resolution capped at ~2K, aspect ratio at 16:9

Ultra-compressed communication mode cutting token usage ~75% while preserving technical accuracy. Supports six intensity levels (lite, full, ultra, wenyan-lite, wenyan-full, wenyan-ultra) triggered by user keywords like "caveman mode," "less tokens," or /caveman command Drops articles, filler words, and hedging while keeping all technical substance and code blocks unchanged; fragments and short synonyms replace verbose phrasing Persists across all responses in a session until explicitly disabled with "stop caveman" or "normal mode" Auto-disables for security warnings, irreversible action confirmations, and multi-step sequences where compression risks ambiguity; resumes after clarity achieved

Route across 10+ video models (HappyHorse, Kling, Seedance, Veo, Wan, Hailuo, Dreamina) for text-to-video, image-to-video, and video extension. Supports text-to-video (t2v), image-to-video (i2v), and Veo's video-extend endpoint with model selection logic optimized for intent (quality tier, multi-shot identity, physics accuracy, audio sync, speed) Includes 10+ production models with documented prompting patterns: HappyHorse 1.0 (Arena #1, in-pass audio), Kling 3.0 (4K, multi-shot), Seedance v2 (multi-modal, cinematic), Veo 3-1 (physics-respecting), Wan 2-7 (audio-driven lip-sync), plus Hailuo, Dreamina, and legacy tiers Each model route ships its exact schema, invoke command, and prompting tips (e.g., "lead with subject and motion verb", "describe audio inline", "Veo respects physics") Invoked via runcomfy run <vendor>/<model>/<endpoint> with JSON input; triggers on "generate video", "make a video

Smart router across 11+ image models for text-to-image and image-to-image generation via RunComfy CLI. Covers FLUX 2 (Klein, Pro, Dev, Flash, Turbo, Max), Google Nano Banana 2/Pro, OpenAI GPT Image 2, ByteDance Seedream 5/4-5 and Dreamina 4-0, Alibaba Qwen Image, Wan 2-7, and Z-Image Turbo — automatically selects the right model for the user's intent (typography precision, photoreal portraits, sub-second iteration, multi-reference styling, open-weights workflows). Supports both text-to-image and image-to-image/edit endpoints with documented prompting patterns and minimal runcomfy run invocation for each model. Triggers on natural language requests: "generate image", "make a picture", "text to image", "image to image", "i2i", or any explicit ask to create or restyle an image. Includes five core t2i routes (FLUX 2 Klein default, GPT Image 2 for in-image text, Nano Banana 2 for speed, Seedream 5 for photoreal, open-weights variants) and three i2i routes (Nano Banana

Unified CLI for running hundreds of RunComfy AI models from the command line. Supports image generation, video creation, editing, face-swap, lip-sync, upscaling, LoRA training, and more through a single runcomfy run <model_id> command with JSON input Handles authentication via interactive login or environment variable, with token stored securely in ~/.config/runcomfy/token.json Offers three polling modes: synchronous (submit, poll, download), fire-and-forget with --no-wait , and status checking via runcomfy status <request_id> Includes JSON output mode for scripting, batch processing, and piping to other tools; exit codes distinguish retryable errors (75) from fatal ones Foundation skill that all other RunComfy intent-routed skills (image-generation, video-generation, face-swap, etc.) dispatch through

Create talking-head and avatar videos by routing audio + image across five specialized models. Routes across OmniHuman (audio-driven full-body avatar), Wan 2-7 with audio_url (open-weights scene + lip-sync), HappyHorse 1.0 (in-pass audio generation from script), Wan 2-2 Animate (stylized character animation), and Seedance v2 Pro (multi-modal cinematic with reference composition) Classifies user intent—pre-recorded audio or script-only, photoreal or stylized, single shot or cinematic—and picks the matching model with documented prompting patterns Each route ships the exact runcomfy run invoke with schema; supports common workflows like multi-language dubbing, UGC product ads, mascot animation, and chaining with image generation Requires runcomfy CLI installed and authenticated via runcomfy login or RUNCOMFY_TOKEN env var

Swap faces into videos or stills by routing to the right model for your intent. Routes across five specialized models: Wan 2-2 Animate for audio-driven video character swap, Kling Motion Control Pro for motion transfer, GPT Image 2 Edit for multi-reference still compositing, Nano Banana Edit for batch identity-preserving swaps, and Flux Kontext for single-image face edits with prose descriptions. Automatically selects the best endpoint based on whether you're working with video or stills, single-shot or batch, motion-preserving or identity-preserving workflows. Invoked via the runcomfy CLI with JSON input specifying reference images, audio, or video URLs; outputs downloaded to a local directory. Requires explicit user consent and operator responsibility — the skill does not gate inputs, and you must refuse requests targeting real people without consent or aiming at defamatory or harmful synthetic media.