Asset Production Pipelines¶

End-to-end documentation for each grit-mcp asset pipeline, from generation through post-processing to S3 delivery.

Architecture Overview¶

All asset pipelines in The Smithy follow a common three-stage pattern:

flowchart LR
    subgraph Generate["1. Generate"]
        PROMPT[Prompt / Source Asset] --> AI[AI Service]
        AI --> RAW[Raw Output]
    end

    subgraph PostProcess["2. Post-Process"]
        RAW --> QA{Quality Gate}
        QA -->|Pass| CLEAN[Cleaned Asset]
        QA -->|Fail| RETRY[Adjust & Retry]
        RETRY --> AI
    end

    subgraph Deliver["3. Deliver"]
        CLEAN --> S3[S3 Upload + Metadata]
        S3 --> REVIEW[Slack Review]
        REVIEW -->|Approved| DONE[Game-Ready Asset]
    end

Each pipeline is implemented as a set of MCP tools exposed by the grit-mcp server. An operator (human or AI agent) orchestrates the tools in sequence, inspecting intermediate results and making creative decisions at each stage.

Pipelines¶

Pipeline	Tools	Status
Audio SFX	`grit_generate_sfx`, `grit_audio_trim`, `grit_audio_normalize`, `grit_audio_loop_check`, `grit_audio_post_process`, `grit_audio_inspect`, `grit_ingest_audio`	Active
Video (Wan2.2 I2V)	`grit_upload_image`, `grit_generate_video`, `grit_check_status`, `grit_ingest_video`	Active
3D Models (Meshy + Blender)	`grit_meshy_text_to_3d`, `grit_meshy_image_to_3d`, `grit_meshy_check`, `grit_meshy_remesh`, `mesh_inspect`, `mesh_decimate`, `mesh_retopo`, `mesh_export_fbx`, `grit_mesh_ingest`	Active
Storyboard Assembly	`grit_list_clips`, `grit_produce_clip`, `grit_assemble_premiere`	Active

Shared Infrastructure¶

All pipelines share the following infrastructure components:

S3 Asset Library¶

Assets are stored in a shared S3 bucket with a consistent path convention per pipeline:

Pipeline	S3 Prefix	Naming Pattern
Audio SFX	`grit/audio/sfx/{group}/{element}/`	`SFX_{GROUP}_{ELEMENT}_{PHASE}.wav`
3D Models	`grit/3d_models/`	Per-model directories
Video	`grit/video/`	Per-generation directories

Every uploaded asset includes a provenance sidecar (.meta.json) containing license, generator, prompt text, and GitLab issue reference. The sidecar ensures traceability from final game asset back to the original creative brief.

S3 Asset Management Tools¶

Three cross-cutting MCP tools manage the asset library independent of any specific pipeline:

Tool	Description
`upload_asset`	Upload a local file to S3 (auto-detects MIME type)
`list_assets`	List/filter assets by prefix and media type (`image`, `video`, `audio`)
`download_asset`	Download an S3 object to a local path

See Grit Tooling — S3 Asset Management for full parameter details.

Slack Review Workflow¶

Before an asset is considered game-ready, it can be posted to Slack for team approval:

Tool	Description
`grit_post_for_review`	Post an asset to Slack for thumbs-up / thumbs-down review
`grit_check_reviews`	Check the current review status of a posted asset

This provides a lightweight creative sign-off loop without leaving the MCP tool workflow.

External Service Dependencies¶

Dependency	Used By	Notes
ElevenLabs API	Audio SFX	Text-to-SFX generation
SoX (Homebrew)	Audio SFX	Trimming with fade-in/out
FFmpeg (Homebrew)	Audio SFX	EBU R128 LUFS normalization
Meshy AI	3D Models	Text/image-to-3D generation
Blender (headless)	3D Models	Inspect, decimate, retopo, FBX export
Tensor.art (Wan2.2)	Video I2V	Image-to-video generation
boto3 / S3	All pipelines	Asset storage and retrieval

Common Workflow Pattern¶

While each pipeline has its own tools and creative decisions, the operator workflow follows a consistent pattern:

Generate — Call the generation tool with a prompt or source asset.
Inspect — Review the raw output (audio properties, mesh poly counts, video quality).
Post-process — Apply pipeline-specific cleanup (trim, normalize, decimate, retopo).
Ingest — Upload the final asset to S3 with provenance metadata.
Review — Post to Slack for team approval.