Vizuara Books · growth engineSample outputs, made from the skill files.
Everything below was produced by the agent skill files you are reviewing, using the master style file and the PaperBanana image pipeline. Real book figures, real generated art, real copy in the house voice. Nothing here is published yet. This is the review board.
01
Brevo email · Book of the Week
by COURIER · EM-2 · master templateThe master template every campaign reuses: 640px, cream page, white content card, mono kicker,
serif italic headline, a framed hand-drawn book figure with its credit line, exactly one burnt-orange
button, and a compliant footer with {{ unsubscribe }}. COURIER fills the params and posts it to Brevo
via POST /v3/smtp/templates. Copy teaches one real thing (the 512x64 query matrix) before it asks for a click.
02
LinkedIn launch post
by INKWELL · house voiceR
Raj Dandekar
Co-founder, Vizuara AI Labs · MIT PhD
We spent a year drawing AI. Today the shelf opens.
Vizuara Books is 42 technical books where every idea gets a hand-drawn figure. Not stock art, not a diagram bolted on at the end. The figure is how the idea is taught.
Take attention, the part of an LLM everyone quotes and few can picture. A 512 token sentence becomes a 512 by 64 query matrix. It meets a 64 by 512 key transpose. The 512 by 512 result is the attention map you have seen in papers. In the book you watch that map get built, one product at a time, with a drawing for each step.
The books run from building an LLM from scratch to 5D parallelism, vision transformers, RAG in production, and reinforcement learning. 1,664 illustrated capsules so far. One login opens the whole shelf.
The whole shelf is free for a month right now, no card.
Read one: books.vizuara.ai/book/build-llm-from-scratch
👍 Like 💬 Comment ↻ Repost ➤ Send
Why this passes the voice gate:
• No em dashes, no double hyphens anywhere.
• No "excited / thrilled / game-changer / dive in".
• Opens with a concrete claim in 8 words, not a rhetorical question.
• Teaches the actual matrix shapes. It is worth saving even if you never click.
• One plain CTA line. 3 lowercase hashtags at the very end.
• The attached image is a Class A real hand-drawn figure from the book (moat), not generated art.
• No em dashes, no double hyphens anywhere.
• No "excited / thrilled / game-changer / dive in".
• Opens with a concrete claim in 8 words, not a rhetorical question.
• Teaches the actual matrix shapes. It is worth saving even if you never click.
• One plain CTA line. 3 lowercase hashtags at the very end.
• The attached image is a Class A real hand-drawn figure from the book (moat), not generated art.
Decision (Raj, Jul 8): posts always lead with the real book figure or a
book-derived collage. No generic generated heroes. The alternate-hero idea is dead.
03
X thread
by INKWELL · posts via TypefullyV
Vizuara Books @vizuara · 1/5
Everyone says "attention is all you need". Almost nobody can draw what attention actually does. We drew it, step by step. A short thread from one chapter of our new illustrated book library.
V
Vizuara Books @vizuara · 2/5
Start with a sentence of 512 tokens, each a vector of length 64. Stack them: a 512 by 64 matrix. Multiply by three learned weight matrices to make queries, keys, and values. Three copies of the sentence, each asking a different question.
▣ figure attached
V
Vizuara Books @vizuara · 3/5
Queries meet keys. A 512 by 64 query matrix times a 64 by 512 key transpose gives a 512 by 512 grid. Cell (i, j) is how much token i should listen to token j. This grid is the attention map printed in every paper.
▣ figure attached
V
Vizuara Books @vizuara · 4/5
Softmax each row so the weights sum to one, then multiply by the values. Every token walks away holding a blend of the whole sentence, weighted by what it cared about. That is one attention head.
▣ figure attached
V
Vizuara Books @vizuara · 5/5
That is 6 matrix multiplications, drawn instead of described. Every book on Vizuara Books is built this way. The whole shelf is free for a month right now: books.vizuara.ai/book/build-llm-from-scratch
04
Editorial poster + page-flip GIF
by INKWELL · FINALIZED direction (replaces rejected watercolor)
Per book, INKWELL now ships both automatically: one poster + one flip GIF, built from that
book's own figures. The rejected watercolor scenes are archived in gen/ and out of every pipeline.
05
YouTube thumbnail
by INKWELL · art + overlaid serif title
How it is built: the generated base art has empty space in the upper-left by design.
A soft paper scrim goes down the left third for legibility, then the title is set live in the site serif
(italic, the house signature), a mono kicker above, a burnt-orange rule, and the logo bottom-right.
Four words maximum. This is the reusable recipe for every book trailer and short thumbnail. When
YouTube OAuth is connected, HERALD uploads the video and sets this as the thumbnail via
thumbnails.set automatically.
What produced this: growth/STYLE.md (visual + voice law), growth/tools/parallel_figures.py (your PaperBanana bulk pipeline),
and the INKWELL + COURIER skill files. Real book figures from bundles/build-llm-from-scratch (115 figures in that one book).
Still to wire before publishing: YouTube OAuth (one click), the 6 Brevo automations (one sitting with you), and your read of the skill files.
Nothing here has been sent or posted. Tell me what to change: voice, crop, colour, layout, which figure. I iterate, then it goes in the queue.
Still to wire before publishing: YouTube OAuth (one click), the 6 Brevo automations (one sitting with you), and your read of the skill files.
Nothing here has been sent or posted. Tell me what to change: voice, crop, colour, layout, which figure. I iterate, then it goes in the queue.