ai-video-producer

Propose an ffmpeg/edit plan to concatenate clips/edited/ into a rough cut in output/drafts/.

Add a recurring character or subject with appearance and voice notes. Saves to characters/.

Produce a script draft from the brief. Saves to scripts/drafts/.

Finalise the locked draft — encode to spec, embed subtitles, write to output/final/.

Scaffold the AI video production workspace (folder skeleton + project CLAUDE.md) into the current directory.

Use this agent to upscale and/or convert aspect ratio for video clips — e.g. taking a 720p 16:9 generation and producing a 1080p 9:16 vertical for Reels/Shorts/TikTok, or upscaling SD-era footage for inclusion alongside HD generations. Picks between resize-only, letterbox, blur-fill, smart-crop, and AI upscaling depending on the source-vs-target relationship and the user's quality bar. GPU-aware: uses NVENC/AMF/QSV when available, falls back to CPU. For AI upscaling defers to the existing `upscale-and-interpolate` skill or external tools (Topaz, Real-ESRGAN) and does not reinvent them. <example> user: "Convert all my 16:9 clips to 9:16 for Shorts, upscale to 1080×1920" assistant: "Launching the aspect-converter agent." </example> <example> user: "This 720p clip needs to sit in a 4K timeline — upscale it" assistant: "I'll use the aspect-converter agent." </example>

Use this agent when the user has produced video elements in parts (clips, segments, scenes, intro/outro, B-roll inserts) and is ready to join and render them into a single deliverable. Detects the host's GPU (NVIDIA / AMD / Intel / CPU-only) and picks the right ffmpeg encoder (NVENC / AMF / QSV / libx264 / libx265). Handles concat-demuxer for same-codec same-resolution sources and re-encode-and-concat for mixed sources. Respects the workspace's `clips/selected/` ordering and writes to `output/`. <example> user: "I'm done picking takes — concat them and render the rough cut" assistant: "Launching the concatenator agent to detect GPU and assemble." </example> <example> user: "Stitch the intro, the three scene clips, and the outro into one MP4" assistant: "I'll use the concatenator agent — it'll pick the right encoder for your GPU." </example>

Use this agent to normalise a set of video clips so they can be safely concatenated or compared — equalises loudness (EBU R128 / -16 LUFS by default), aligns sample rate / channel layout, conforms framerate and pixel format, and optionally applies basic colour-level normalisation (full↔limited range, BT.601↔BT.709). Run before `concatenator` when sources are mixed (different generation models, screen recordings, ElevenLabs voice tracks, etc.). Writes normalised copies to `clips/normalised/` — never overwrites originals. <example> user: "These clips were generated with different models and the audio levels are all over the place — fix them" assistant: "Launching the normalizer agent." </example>

Use this agent to turn an approved pipeline SPEC.md (produced by pipeline-scaffolder) into an executable pipeline — runnable scripts/configs that `/run-pipeline` can invoke. Generates per-stage runner files, parameter templates, an entry-point script, and a pipeline README. Wires up calls to fal/replicate (or other providers) using the project's MCP servers and respects the workspace's path conventions. Use after the user approves a SPEC.md. <example> user: "Build out the pipeline I just scaffolded — talking-head v1" assistant: "Launching the pipeline-builder agent to generate the runners from SPEC.md." </example>

Use this agent to set up a private GitHub repository where the user versions reusable AI video pipelines across projects. Creates the repo via `gh`, scaffolds an opinionated structure (one folder per pipeline, shared model/cost reference, CHANGELOG, README), seeds it from any pipelines already present in the current AI-Video-Producer workspace, and wires the workspace to consume pipelines from there. Use the first time the user wants pipelines to live outside a single video project, or when migrating an existing pipeline collection into version control. <example> user: "Set me up a private repo for my pipelines" assistant: "I'll launch the pipeline-repo-setup agent to create and seed it." </example>

Estimate API costs for a planned pipeline or for an entire script before any generation runs. Reads pipeline SPEC.md(s), the storyboard, and `brief/tools-and-models.md`, then projects cost per shot and total — with a low / typical / high range that accounts for take iteration. Writes a budget report to `budgets/estimate-<YYYY-MM-DD>.md`. Use before `/run-pipeline` on any non-trivial project.

Export an AI-Video-Producer pipeline to ComfyUI workflow JSON — produce a workflow the user can load in ComfyUI to reproduce (or approximate) a hosted-API pipeline locally. Best-effort: maps fal/replicate model calls to equivalent ComfyUI checkpoints/LoRAs/samplers and flags stages that have no native ComfyUI equivalent. Use when the user wants to take a pipeline that's been working via APIs and bring it onto local hardware (cost reduction, offline work, fine-grained control).

Import a ComfyUI workflow JSON (the API-format export from ComfyUI's "Save (API Format)" option, or the UI-format graph JSON) and turn it into an AI-Video-Producer pipeline SPEC plus stage runners. Maps Comfy nodes to pipeline stages, surfaces the model checkpoints and LoRAs needed, and flags nodes that have no clean equivalent in a hosted-API pipeline. Use when the user has a ComfyUI workflow that works locally and wants to either run it inside a project workspace or document it as a versioned pipeline.

Research what fal.ai and Replicate (and optionally other providers — Runway, Kling, Pika, Hedra, ElevenLabs, OpenAI) currently offer for a given workload — text-to-image, text-to-video, image-to-video, lip-sync, voice, upscaling, interpolation — and recommend the best-fit model for the project's brief. Compares quality reputation, price, max duration/resolution, aspect ratio support, and known failure modes. Updates `brief/tools-and-models.md` with the recommendation and rationale on user approval.

Intelligent, preference-driven model recommendation across fal.ai, Replicate, WaveSpeedAI, and MiniMax (Hailuo). Asks the user a short set of preference questions (workload, priority — quality vs speed vs cost, max budget per output, resolution/aspect/duration, NSFW tolerance, must-have features like lip-sync or audio), queries each available provider, and returns a ranked shortlist (3–5 options) with approximate per-output costs, quality/speed notes, and a recommendation. Differs from `model-researcher`: live-API backed, cost-first, and conversational about preferences.