Designing AI‑Driven Content Layouts for GUTSTRING: A System That Reads the Art
As we continue to build our platform, we came across an idea to implement a showcase section for highlighted creators. The content they submit ranges from short text snippets to several-minute videos, image series, and audio pieces.
Our question was simple but ambitious: instead of hand-designing a unique page for every artist, could software understand the work and choose a presentation that feels intentional — without losing the brutalist, editorial tone of GUTSTRING?
SIGN UP FOR EARLY ACCESS NOW AND RECEIVE A LIMITED TIME ONLY PROFILE BADGE.
How Might this Work?
Squarespace AI Assistant
Squarespace’s AI assistant is a useful reference. You can start from a blank page, add minimal copy and a few images, and receive several AI-generated layout suggestions on the fly. The templates are often generic, but the premise is solid: infer structure from content, not the other way around. For GUTSTRING, we want something stricter and more artistic — layouts that feel like a gallery or poetry broadside, not a SaaS landing page.
Fundamentally, this is about reading what the artist submitted and finding a restrained, repeatable way to present it.
Our Method
Storing the Content
We assume content already exists in Azure Blob Storage — one folder per artist, all files at rest under a predictable prefix (e.g. gutstring/showcase/Nessuno/). Blob storage gives us file-type flexibility (JPG, PNG, TXT, MP4, MP3, etc.) and a single source of truth without coupling showcase to our main API or Cosmos DB. Discover remains the live, social feed; showcase is a curated exhibition, generated offline and served as static layout JSON.
Understanding Content Properties (Without Vision, for Now)
Before any “design agent” runs, a compiler pass inventories each folder:
File type — extension and MIME (text, image, video, audio, other)
Metadata — size, last modified (order and sanity checks)
Text — full body ingested at build time for .txt / .md files (paragraph splits on blank lines)
Images — dimensions via Pillow (aspect ratio for grid placement later)
Video / audio — duration via ffprobe when we enable it
We are not using vision models in the MVP. Mood, colour, and theme tags can wait; the first version is deterministic and auditable.
A Design Agent
The long-term design agent will choose among a small set of layout templates and theme tokens (alignment, spacing, accent colour) — not free-form CSS. Actions might include: pick hero-media vs image-grid, set text to left or centre, reorder sections. Outputs must validate against a JSON schema so bad generations never reach production.
A Hard Rule: Use Every File
Every blob in an artist folder must appear exactly once in the generated layout. The compiler asserts usedBlobPaths === inventory before publish. If a PDF or odd file appears, it gets a file-link block rather than being dropped. This keeps the exhibition honest: the page is the folder, not a curator’s subset.
Our MVP
For the first shipped version, we deliberately did not wait for model tuning or a fully autonomous agent. Instead we built a rule-based layout compiler — a one-time (or on-demand) script that lists each artist’s blobs, classifies them, emits a versioned layout, and uploads a root showcase-index.json so the home of the showcase section can list every participant with a single fetch.
Layout rules the compiler applies today
Mixed text + media — First image (or video if no image) becomes a wide hero with grayscale/contrast treatment; every text file becomes its own text-block, left-aligned when media is present (poetry convention). Remaining images form a grid (two columns when there are several) or single media-block sections; each video/audio file gets a dedicated block so nothing is skipped.
Text only — All text files are rendered as centered blocks, one after another — appropriate when a portfolio is purely literary.
Images only — No hero preference beyond the first file; all images are placed in an image-grid with aspect-ratio-aware slots so the page doesn’t feel like a single stock banner.
Coverage check — If compile_showcase.py finishes without assigning every path, the build fails loudly with a list of orphan files rather than publishing a partial exhibition.
