Skip to content

Brand assets: how each one was made

Every image in this repository is reproducible from the prompt or the source below. Nothing here is decoration nobody can regenerate — a logo whose prompt was lost is a logo you cannot iterate on.

Choosing the tool before writing the prompt

This is the part that saves an afternoon. The three jobs look similar and need different tools.

You want Use Why
A character (mascot, avatar, hero art) nano_banana_pro on Higgsfield Image models are excellent at this and nothing else competes on time-to-first-good-result.
A diagram whose labels must be correct Excalidraw or Mermaid An image model draws a picture of a diagram. It will misspell a label eventually, and you cannot fix one word without regenerating the whole thing.
A diagram for a landing page, look over precision nano_banana_pro, then verify every word Good enough when the text is four short labels, and it is the look people associate with a friendly open-source project.
An explainer video Remotion Code-driven, so the numbers are the real numbers and a rerun after a product change costs nothing. Akaion already runs it.
Atmosphere — a hero clip, social b-roll Veo 3.1 Beautiful motion, no control over text. Never put a label in it.

The rule that matters: never ship a diagram from an image model without reading every word in it out loud. Models render plausible text — RESTRICED, MODELO FRONTIER, a repeated arrow label — and the errors are invisible to whoever wrote the prompt because they read what they meant.

Practical division of labour, and what this project does:

  • Mermaid in docs/design/hld.md — eight diagrams, versioned with the code, render on GitHub and in the docs site. These are the true ones.
  • An illustrated one for the landing page and for a slide, generated below. It carries four labels and a mascot, which is the ceiling of what an image model should be trusted with.

The mascot

annona-mascot.png — generated with nano_banana_pro (1:1), then background removed with the same provider's remove_background, then resized to 512 px.

3D rendered cute chibi mascot character logo: a friendly Roman praetorian guard
standing centered, glossy Pixar-style rendering, big expressive eyes, warm
smile, wearing a burnished bronze legionary helmet with a deep red horsehair
crest and a verdigris-patinated cuirass with subtle gold trim, holding a small
round shield engraved with a golden wheat sheaf in one hand and a rolled sealed
scroll in the other, a small terracotta amphora spilling golden grain at his
feet, soft studio lighting with a subtle teal rim light, clean solid dark
background, centered composition, highly detailed, appealing toy-like
proportions, brand mascot illustration

Why those objects: the wheat sheaf is the annona, the sealed scroll is the ledger, the amphora is the supply. The verdigris armour is the accent colour used across the docs (#1F6F68). Ask for a dark solid background even when you want transparency — a cut-out from a clean background is far better than one from a busy scene.

The "how it works" illustration

Template. Replace the bracketed parts; keep the structure, because the structure is what makes the model place things predictably.

Hand-drawn whiteboard-style technical infographic, black marker on off-white
paper, clean and friendly, minimal text. Title at top in hand-lettered capitals:
"[TITLE, 2-4 WORDS]".

LEFT COLUMN: [two documents, one with a red padlock labelled "[LABEL 1]", one
plain labelled "[LABEL 2]"].

CENTER: a rounded rectangle box with a soft teal border containing [the mascot,
described in one sentence]. Under the box, hand-lettered: "[PRODUCT]" and
smaller beneath it "[three verbs separated by dots]".

RIGHT COLUMN: three destination boxes stacked vertically, labelled "[DEST 1]",
"[DEST 2]", "[DEST 3]".

ARROWS: a thick green arrow from [source] into the box, then to [allowed
destination]. A red dashed arrow from the box toward [forbidden destination],
crossed with a large red X and a small hand-lettered tag reading "[REFUSAL
WORD]". A separate green arrow from [other source] through the box to [other
destination] with a green check mark.

BOTTOM STRIP: [one visual element] with the hand-lettered caption "[CAPTION]".

Style: excalidraw / notebook sketch aesthetic, slightly wobbly lines, generous
white space, only the words listed above, no other text, no paragraphs, no
watermark, no logos.

Five rules learned the expensive way:

  1. Name the layout in capitals (LEFT COLUMN, CENTER, ARROWS). Models follow positional instructions far better than they follow prose.
  2. Cap every label at three words. Long strings are where the spelling goes.
  3. Say "only the words listed above" or the model will invent captions, and the invented ones are always the ones that end up wrong.
  4. Ask for generous white space. Crowded sketches read as noise at the size a landing page uses.
  5. Use 16:9 for anything going above the fold on a page or into a slide.

Verification checklist before shipping one

  • [ ] every word matches what you asked for, character by character
  • [ ] the arrows point the way the product actually behaves
  • [ ] nothing implies a guarantee the code does not make
  • [ ] no invented vendor logo (a model will draw an approximate one, and that is somebody else's trademark, approximately)

What to use for a video

Remotion, not a video model, for anything explaining the product: it is React that renders frames, so the numbers on screen are read from the same place the README reads them, and a rerun after a change costs a build. Akaion already has akaion-app-remotion.

A generative video model — Veo 3.1, or Higgsfield's catalogue — earns its place in exactly one role: a few seconds of atmosphere behind a title card. The moment a label has to be legible, it is the wrong tool, and no amount of prompting changes that.