From a scribble
to a finished leaf.

Every writer knows the friction: an idea arrives as a napkin scribble, a voice note, or three messy bullets — and dies inside a notes app. Blogging tools assume you already have a finished draft. Social platforms reward speed over depth. Neither respects the middle: the messy, private space where thought becomes form.

We watched a dozen writers try to move a raw idea into a public post. On average they abandoned three drafts before shipping one. The drop-off was not skill — it was the cost of turning a fragment into structure.

Suvadi (சுவடி) is the Tamil word for a palm-leaf manuscript — a personal drawer of notes, essays, and scribbles bound with string. Before print, before feeds, this was how knowledge travelled: slowly, tactilely, in objects you could hold.

We took the metaphor literally. Every writer gets a cabinet of drawers. Every article is a leaf. Readers pull a drawer, hold a leaf, leave a stamp. The interface behaves like an object, not a stream.

A cabinet you can finish reading beats a feed that never ends.

The core loop is five steps. Each removes a decision the writer would otherwise make alone. The AI drafts; the writer edits. The tool never ships anything without a human pass.

Under the hood the pipeline is boring on purpose. Gemini 3 Flash returns a strict JSON structure — title, hero-illustration subject, sections. Nano Banana adds sparse line-art only where the model flagged it worth illustrating.

A Suvadi article is a JSONB array of blocks. Each block knows how to render itself, how to be edited, how to be dragged, and how to feed the share generator. Hover a tile to see what it does.

The block system is what lets the same article become a printable paper, a vertical story, and a carousel deck without re-writing a word. Structure is the product.

We rejected the default AI aesthetic — purple gradients, glassy cards, generic sans-serifs. Suvadi is high-contrast black on white, sharp borders, no shadows. Special Elite for display. Courier Prime for captions. Every hover is a color inversion, not a drop-shadow.

The visual language says: this is a workshop, not a feed. Line-art illustrations replace stock photos. Palm-leaf dividers replace horizontal rules. A kolam loader replaces the spinner.

Discovery lives at /feed as "The Square" — an asymmetric bento mosaic of today's leaves. Tile sizes vary deterministically by article length and stamp count.

No infinite scroll. No algorithmic ranking. When you reach the bottom you're done — and that is the feature.

One leaf, three shapes. The share modal reads the block array and renders three exports: Paper, Story, and Carousel — a curated .zip of 6–9 slides assembled from the same block data.

The carousel is the moat. Every competitor makes you screenshot. We treat the deck as a first-class output of the article — with TL;DR, deep-dive sections, a curiosity teaser, and an outro, deduped by Jaccard similarity.

Depth doesn't have to mean friction. When the AI does the scaffolding, the writer keeps the taste — and the whole loop takes under a minute. The trick was refusing to let the AI ship. It drafts. You publish.

The unexpected win was the share pipeline. Because every article is structured data, we can generate a whole carousel from one block array. New export formats — slides, newsletters, audio — are a renderer away.

Structure is the product. Everything else is a renderer.