Skills
96,981 packaged capabilities and playbooks for agents. Install any of them with one command.
Creative & media
Generate and edit images using Google's Nano Banana 2 (Gemini 3.1 Flash Image Preview) API. This skill should be used when the user asks to create or modify images, especially when they need fast iteration, explicit aspect-ratio control, or resolution control from 512px to 4K.
Use when the user asks to generate, edit, extend, poll, list, download, or delete Sora videos, create reusable non-human Sora character references, or run local multi-video queues via the bundled CLI (`scripts/sora.py`); includes requests like: (i) generate AI video, (ii) edit this Sora clip, (iii) extend this video, (iv) create a character reference, (v) download video/thumbnail/spritesheet, and (vi) Sora batch planning; requires `OPENAI_API_KEY` and Sora API access.
Generate and conversationally edit short videos with stable Google Gemini Omni Flash (`gemini-omni-1.1-flash`). Use when: (1) iterating on a clip with natural-language edits instead of regenerating ("make the phone invisible, keep everything else the same"), (2) generating 3-10s 720p clips with synthesized audio, rendered on-screen text, or timecoded beats, (3) binding reference images to roles with <FIRST_FRAME>/<IMAGE_REF_N> prompt tags, (4) editing an existing uploaded video. Accessed via the `gemini_omni_video` tool using the project's GEMINI_API_KEY/GOOGLE_API_KEY — the same key as Imagen and Google TTS
Generate configurable chat motion overlays from a transcript or screenshot, including plain bubble scenes, app-style chat containers, optional device frames, preset or uploaded avatars, nickname display rules, and transparent-video-ready Remotion bundles. Use when Codex needs to create reusable short-form chat clips for Douyin, demo videos, story reenactments, social-message proof scenes, or embeddable alpha overlays for Hyperframe and Remotion workflows.
Generate a swinging spotlight text-reveal HTML animation with configurable text, swing angle, lamp scale, glow, and colors. Use when users ask for 聚光灯扫字动画, spotlight text reveal, light logo reveal, 发光文字揭示动画, or want a reusable HTML animation instead of a static image.
Generate high-quality images using the Z-Image API from ModelScope. Use this skill when the user wants to generate images with the specific Z-Image model or ModelScope API they provided. Trigger words: 'Zimage', 'ModelScope', 'generate zimage'.
Build Blender mesh surfaces directly from extracted 2D contours/masks instead of approximate primitives. Use for 1:1 mascot/logo reconstruction, exact leaf/petal silhouettes, filled wireframe shapes, shallow bas-relief forms, or when the model outline must match a front template before adding depth.
Register front, side, back, and top orthographic reference views into a shared Blender coordinate contract. Use when a 3D model must match multi-view wireframes/templates, when side/back/top silhouettes are off, or before adding depth to a front-locked contour model.
You MUST use this when drafting, rewriting, editing, or reviewing English or Ukrainian prose for clarity, concision, tone, naturalness, or audience fit, including documentation, messages, articles, and product copy. Not for translation, agent instructions, or UI design. Specialized dash and negative-parallelism audits belong to dashfix and negafix; validation of technical review findings belongs to review-resolution.
Convert 2D orthographic wireframe PNG drawings into 3D Blender models exported as glTF/GLB. Use this skill whenever the user provides wireframe images (technical drawings, line drawings, orthographic views, side/front/back panels) and wants to generate a 3D model, mesh, or .glb file. Triggers on phrases like "convert this wireframe to 3D", "make a 3D model from these drawings", "build a model from this wireframe", "generate GLB from these views", or any image-to-3D-mesh request involving line drawings. Make sure to use this skill even if the user does not explicitly say "wireframe" — it also covers "orthographic views", "technical drawings", "line drawings of objects", "front and side views". Requires the Blender MCP addon to be running (port 9876) and Python with opencv-python, numpy, scipy installed.
Closed-loop compare-adjust-repeat workflow for fitting Blender models to supplied front/side/back/top templates and originals. Use when the user asks to compare the product to templates/originals and adjust until it fits across all dimensions, or when all views must pass measurable bbox/centroid/silhouette/edge validation before export.
Polishes and optimizes novel content. Applies when users ask things like "polish this novel for me", "improve the writing style", "optimize the chapter pacing", "strengthen this exciting beat", "make the dialogue more natural", "make this part more satisfying", "polish the novel's prose", "adjust the chapter pacing", "make the dialogue more realistic", "revise this passage for me", "polish the novel", "optimize the exciting beats" or "make this part more immersive". Provides 3 levels of polishing, focusing on writing style and content, with targeted improvements such as style adaptation, tightened pacing, stronger exciting beats and dialogue optimization. **The polished result directly modifies the chapters/ directory, and is automatically backed up to .sumeru/write/original/ before modification.****When polishing in bulk, sub-agents are used to process in parallel, with each agent handling up to 3 chapters.**
Comic drama generator - a tool for generating manga-style short dramas based on Seedance. Supports automatically generating comic drama storyboard scripts from a protagonist image and producing videos. Suitable for creating manga-style short videos, character stories, animated shorts, etc. Use this skill when the user wants to generate manga-style video shorts, character stories, or comic dramas.
This skill should be used when the user asks to "make a Vox-style map animation", "animate a map zoom", "draw a route on a map", "highlight a country/region", "use Google Earth Studio", "drop pins on a map and animate them", "pan/orbit a camera over a map", or "add map graphics to an explainer video". Covers Earth Studio→After Effects camera import, GeoJSON/SVG vector map motion, pins/routes/highlights/labels, the 12fps stutter cadence, and data-driven Remotion renders.
The CapCut craft skill — edit short-form social video (TikTok, Reels, Shorts) fast and safely: retention-paced cuts, auto-captions, beat-synced sound, exports. Use when someone wants to edit in CapCut, cut a reel/short, set up auto-captions, use CapCut's AI tools, pick a plan, or asks whether CapCut music/assets or client work are safe commercially. Uses the EDITS framework. Reads short-form-video-script + brand-profile/design-and-templates + voice-builder first. The agent plans the edit and flags the licensing/ToS gates; the HUMAN executes in CapCut and approves; WoopSocial publishes the finished export. Safety spine: client/NDA work edits locally, never synced (the ToS licenses uploads), and music licensing has fine print — never pirated assets, non-consensual lip-sync, or uncorrected captions. Distinct from short-form-video-script (the script), captions-and-clipping/opus-clip (clipping long video), descript (long-form), canva (graphics), and ai-video/veo-3/kling/runway (generation).
Run a self-refinement loop when a Blender result is subpar, user expectations are not met, validation fails, or repeated issues reveal missing skill depth. The loop diagnoses the gap, decides whether to repair the artifact or acquire/refine a generic skill, sanitizes lessons for reusable publication, updates docs/versioning, validates, and prepares stage/commit/push when explicitly requested.
Use whenever you build, modify, or debug an application integration with Runway Dev: inspect the app, connect Dev MCP when useful, find current API contracts, integrate through the official SDK, and keep credentials server-side. Load the matching surface skill for models, Model Routers, Characters, Recipes, or published Workflows.
Novel outline, worldbuilding, and character design. For requests such as "help me write a novel outline", "design the main character", "build the world setting", "structure the plot framework", "write a detailed volume outline", "design novel characters", "create a xianxia world", "map out the plot", "character settings", "chapter-by-chapter outline", "plan the payoff moments", "make a character sheet", "build the novel's world", it generates complete world settings, character cards, plot outlines, and payoff plans, with automatic compliance checks to avoid infringement risks. When generating large numbers of chapter outlines, sub-agents are used in parallel, with each agent handling outlines for at most 3 chapters
Record an autonomous 50 FPS promo video of your game for social media. Use when the user says "record a video", "make a promo", "capture gameplay", "record gameplay footage", or "make a trailer for my game". Do NOT use for gameplay testing (use qa-game).
Create novel chapter content. Applies to requests such as "write me a chapter of a novel", "continue with what comes next", "generate an XX plot point", "batch write web-novel chapters", "expand/rewrite this passage", "write me an XX plot point", "continue the novel", "expand this passage", "rewrite this chapter", "batch generate novel chapters", "write an opening chapter", "write a climax scene", "generate novel content", "help me write novel content". Supports single-chapter and multi-chapter batch generation, continuation, rewriting and expansion modes, automatically matches web-novel pacing, keeps characters and plot consistent, and **uses sub-agents in parallel during batch generation, with each agent handling at most 3 chapters**.
Clone or update https://github.com/sxhzju/wechat-2d and render the default WeChat-style 2D chat motion video with Remotion. Use when users ask for 微信聊天动画, wechat 2d chat render, 微信视频消息动效, or exporting the default demo from the wechat-2d project.
Create professional cinematic scripts for AI video generation with character consistency and cinematography knowledge. Use when the user wants to write a cinematic script, create story contexts with characters, generate image prompts for AI video tools (Midjourney, Sora, Veo), or needs cinematography guidance (camera angles, lighting, color grading). Also use for character consistency sheets, voice profiles, anachronism detection, and saving scripts to Google Drive.
Analyze short-form videos with Gemini AI to extract hooks, content structure, and replicable patterns. Supports Instagram Reels, TikTok, and YouTube Shorts. Use when asked to: - Analyze video content for hooks and structure - Extract replicable formulas from viral videos - Understand why a video performed well - Get AI analysis of video content patterns Triggers: "analyze videos", "extract hooks", "video analysis", "analyze reels", "what makes this video work", "hook analysis", "content structure analysis"
Novel logic and plot review. Applies when users ask things like “check my novel for bugs”, “see if the timeline is contradictory”, “check for out-of-character characters”, “find plot conflicts”, “see whether all foreshadowing is resolved”, “check whether the plot makes sense”, “look for plot holes”, “characters behaving out of character”, “check whether the timeline is consistent”, “find contradictions in the novel”, “help me map out all the foreshadowing”, “plot bug check for my novel”, “logic hole investigation”. It detects timeline conflicts, logic holes, out-of-character characters and unresolved foreshadowing, enforces word-count checks so chapters meet the required length, generates a detailed issue list and fixes all problems automatically. Fixes are written directly into the chapters/ directory, with an automatic backup to .sumeru/write/original/ before modifying. When severe problems are found it automatically revises the outline and rewrites the affected chapters. A zero-tolerance mechanism ensures all fatal problems must be resolved. In batch reviews, sub-agents work in parallel, each handling at most 3 chapters.
Record a microphone or edit video/audio with the auto-editor CLI — cut out dead space (silence, motionlessness, black frames), set the pace, make manual cuts, and render the result. Use when someone wants to capture and trim a microphone, remove silence, shorten a recording, speed through quiet parts, or otherwise edit media with auto-editor.
Topic planning for web novels and fiction. Applies when users say things like "I want to write a novel but don't know what", "help me come up with a novel theme", "find web novel ideas", "analyze which themes are hot", "evaluate a novel topic", "which web novel themes sell well now", "give me a few novel ideas", "invent a novel cheat ability", "help me find a popular theme", "which themes are easy to make popular", "web novel market trend analysis", "help me pick a novel topic". Produces multiple topic proposals and market analyses covering cheat-ability design, core selling points, satisfaction patterns, and feasibility assessment.
Use to write a TikTok script — a native, sound-aware short-video script optimized for TikTok's discovery engine (the FYP), built around a brutal 1-second hook, watch-time, and rewatches. Run when the user says "TikTok script," "TikTok video," "script for TikTok," "make a TikTok," "storytime," "TikTok hook," or wants short-form video for TikTok specifically. Reads brand-profile and voice first; rides a trending sound/format where it genuinely fits (live trend research goes to trend-jacking); structures for completion + rewatch; optimizes for TikTok search; outputs a production doc (shot - on-screen text - spoken - sound - timing) plus a Veo prompt pack for faceless/b-roll shots (native filming often beats AI on TikTok). For Instagram Reels use reels-script; for YouTube Shorts use youtube-shorts; for the platform-agnostic scripting craft use short-form-video-script; hands caption to caption-writer and scheduling to scheduling-and-queue.
Transcribe audio to text using Azure AI Speech (Fast Transcription REST API). Use when converting audio/video to text, generating subtitles, or processing spoken content in OpenMontage. Optional cloud STT provider — preferred when AZURE_SPEECH_KEY is configured; the local faster-whisper `transcriber` is the default offline path
Generate photorealistic 3D worlds and environments with the World Labs Marble API — Gaussian Splat scenes from text prompts or reference images. Use when the user says "generate a 3D world", "create an environment", "make a 3D scene", or "use World Labs". Requires WLT_API_KEY environment variable.
End-to-end TikTok or Instagram slideshow production with vibe-driven prompt writing, local JSON slide management, optional localhost review GUI, batch GPT Image 2 generation, and text overlay compositing. Use this when the user wants to create slideshows or carousels from scratch, including script planning, image prompts, and export-ready slide sequences.
Detect texture-atlas regions and map each source part to its own UV rectangle or projected UV surface. Use when a texture pack must fit a Blender model 1:1, when textures look off/stretched, when alpha decals or lightmaps are provided, or before exporting a textured GLB for a brand mascot/logo.
Language-specific guidelines for producing high-quality Traditional Chinese (zh-TW) content: script, full-width punctuation, spacing, forms of address, grammar constraints, and Taiwan terminology. Includes the no-calque rule against literal translations of English phrasal verbs such as 折入 for fold in. Use when writing any kind of Chinese content. Including blog posts, notes, technical articles, technical writing, chitchat, social media posts, etc., even when you are just sending a text message. Also use when reviewing or editing existing Chinese content for language and terminology compliance. Always apply `writing-style-guideline` alongside this skill for structure, tone, rhetorical constraints, and banned phrases.
End-to-end web/novel world-building master for serialized fiction. Fits requests such as "I want to write a novel", "write me a web novel", "create a novel from scratch", "help me write an XX-type novel", "I want to write a book", "give me the whole novel-writing process", "write a novel automatically", "one-stop novel writing service", "help me finish a whole novel", "I only have an idea, how do I write a novel", and other overall novel-writing needs. It automatically coordinates topic selection, outlining, writing, review, and polishing across the full workflow, so even a rough idea can become a complete work from 0 to 1 with a unified story world. All chapter-level operations are handled in parallel by sub-agents, with each agent responsible for at most 3 chapters.
Generate and validate music with Google Lyria 3 through the Gemini Interactions API. Use before calling OpenMontage `google_music`, designing Lyria 3 Clip or Pro prompts, using image-to-music or custom lyrics, choosing between Lyria 3 and Lyria RealTime, diagnosing Google music-generation failures, or preparing exact-duration music for a video.
This skill should be used when the user asks to "create a plot arc", "story structure", "add a plot point", "story timeline", "track foreshadowing", "pacing", "sagging middle", "act structure", "story arc", "plot outline", "snowflake method", or wants to plan and manage the narrative structure of a story. It owns book-level pacing; NOT for the craft of scene outcomes or chapter hooks, which this skill only plans in the outline (use scene-craft), or a pacing check as a revision pass (use revision-continuity).
Build explicit learn/do-not-copy contracts for image and video generation references. Use this when a prompt uses benchmark videos, contact sheets, frames, or product images and you need to state exactly what the model should learn, what identity elements must change, and which references should be excluded from the first test.
ElevenLabs text-to-speech via inference.sh: Eleven v4 and v4 Turbo (newest, most expressive, audio tags, realtime streaming), Eleven v3, Multilingual v2 and Flash v2.5, each as its own app with 22+ premium voices. Capabilities: text-to-speech, realtime text-to-speech streaming, audio tags for emotion and delivery, voice selection, stability control, custom and cloned voices. Use for: voiceovers, audiobooks, video narration, podcasts, voice agents, accessibility, IVR. Triggers: elevenlabs, eleven labs, eleven v4, eleven v4 turbo, eleven v3, elevenlabs tts, premium tts, realtime tts, streaming tts, professional voice, ai voice, high quality tts, multilingual tts, voice generation, natural speech, realistic voice, voice over, speech synthesis
Turn multiview fit reports into a sequential or parallel repair queue for aligning Blender products to source-of-truth templates. Use after validation shows disalignment, when the agent must iteratively fix wireframe, texture, lighting, or projection mismatches, and when skill gaps should trigger self-refinement before another rebuild.
The growth engine for YouTube Shorts in 2026. Use when someone asks to "grow on Shorts," "get views on YouTube Shorts," "write a Shorts script," "why did my Short flop," "make a vertical YouTube video," or wants a Shorts strategy. Drafts the script; the human films and edits; publishing routes through scheduling-and-queue -> WoopSocial; analytics are read natively in YouTube Studio. Fully decoupled from, and a sibling to, youtube-long-form.
Rewrites Russian text with signs of AI generation so it reads as if written by a real person: removes bureaucratic language, ChatGPT clichés, empty lead-ins, “not just X, but Y,” triples, excess dashes and bold, and junk chat insertions (contentReference, utm_source, invisible characters). Preserves and checks numbers, names, dates, conditions, and negations. Use when asked to “humanize,” “rewrite naturally,” “remove GPT-ness,” “make it lively,” “remove bureaucratic language / fluff / clichés,” when text “sounds like AI,” to “edit” or “polish” text, and also for analysis without editing: “what gives away AI,” “check the text.” Russian only.
Review storyboard prompts and video-request prompts before generation. Use this when a prompt draft already exists and you need to catch weak first frames, drift risk, missing constraints, bad product timing, or generic ad-like language before spending model credits.
All-around Feishu whiteboard operations · Choose any of 5 drawing paths: (A) Mermaid/PlantUML server-side rendering (mind maps/sequence diagrams/class diagrams/pie charts/flowcharts/Gantt charts, the whole diagram as one node, Feishu handles the layout automatically) (B) Mermaid local engine whiteboard-cli (bypasses server-side limits such as par/participant count, every node is clickable and editable) (C) Free-form SVG drawing (any visual: flywheel/fishbone/Dashboard/poster/illustration/architecture diagram, etc., each SVG element → 1 native node that can be edited individually) → use scripts/svg_to_board.py to translate SVG into native Feishu nodes (D) Simple SVG single-node decoration (icons/stamps/small elements, SVG within 2KB) (E) Finely laid-out architecture diagrams (hand-written node JSON, absolute coordinates + color scheme + connector ID references) Use when the user asks to "draw/diagram/whiteboard/draw an architecture diagram/draw a flowchart/draw a flywheel/draw a fishbone/draw a roadmap/draw a Dashboard/draw an illustration/draw a poster/AI free-form drawing/put SVG on the board/clone a board/upload an image to the board/visualization/node diagram/refined layout". In particular, when users report "half a building in the bottom-right corner", "z_index messed up", "duplicated nodes", "complex diagram not fully rendered", "mermaid server-side failure", be sure to read references/pitfalls.md for troubleshooting. Write operations (add-board/create-notes/update/delete/clone/svg-import/upload-image/import) use the App Token (Bot identity by default) with no login needed; even if the user has run auth login, write commands still stay on the App identity. Read operations (image/nodes/export-code/svg-export/lint) automatically use the User Token after login and fall back to the App Token when not logged in. svg-export exports an SVG visual snapshot of the entire board (POST /export, base64 decoded), valid for any board (not limited to SVG nodes); combined with svg_to_board.py / board import it enables an "export → edit → write back" loop.
Turn approved storyboard logic, beat sheets, or prompt plans into provider-ready short-form video requests. Use this when the segment structure is already known and you need a model-agnostic request architecture that can later map cleanly into Seedance or other video generators.
Download videos from YouTube, Bilibili, Twitter, and thousands of other sites using yt-dlp. Use when the user provides a video URL and wants to download it, extract audio (MP3), download subtitles, or select video quality. Triggers on phrases like "下载视频", "download video", "yt-dlp", "YouTube", "B站", "抖音", "提取音频", "extract audio".
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.