Amini Illustrations

---
name: amini-illustrations
description: "Generate on-brand Amini editorial illustrations for articles, decks, docs, blog posts, product pages, LinkedIn posts and investor material. Use for requests like article illustration, hero image, explainer graphic, shot list, illustration strategy, diagram, concept art, mascot, or Ami. Dark-first Space Black ground, Stone Grey line art, single Sunglow accent, deadpan Ami character performing the core action. Includes a palette auditor that enforces the Amini brand hex."
license: MIT
metadata:
  author: Amini
  version: "1.0.0"
---

# Amini Illustrations

Design and generate 16:9 editorial illustrations for Amini content — articles, decks, docs, product pages, social posts, investor and government material.

The goal is not decoration, marketing gloss, or a technical diagram. It is to take one judgement, structure, state, or metaphor from the piece and render it as a single restrained image that a minister, an investor, or an engineer would all read the same way.

The recurring character is **Ami**: a solid charcoal figure with blank white dot eyes and thin legs, operating infrastructure with a completely level expression. Ami is not a mascot. Ami must perform the core action of the image — if the idea survives with Ami removed, Ami is decoration and the prompt needs rewriting.

## Read these as needed

Load only what the task requires:

- `references/style-dna.md` — palette roles, line language, the restraint rules, hard prohibitions.
- `references/ami-ip.md` — Ami's form, register, action vocabulary, and prohibitions.
- `references/composition-patterns.md` — structure types, Amini's subject domain, and the metaphor engine.
- `references/prompt-template.md` — the per-image generation prompt and edit prompts.
- `references/qa-checklist.md` — what to check after generation and how to iterate.
- `assets/ami-reference/` — the canonical Ami, dark and light. Style calibration only. Do not reproduce this pose or subject.
- `assets/examples/01-digitisation-flow.png` — one generated-and-audited image showing the target look. Calibration for palette, line weight, and emptiness. Do not reuse its composition.
- `examples/6g-isac-illustrations/` — a worked shot list for a research paper, plus a real palette audit run. Shows the planning and QA artefacts, not the images.

## Brand ground rules

These come from the Amini design system and are not negotiable:

- **Dark is the default.** Space Black `#121212` ground unless light mode is explicitly requested.
- **Sunglow `#FFC83C` is the only accent** and never a background fill. One accent per image, marking the one thing that matters.
- **Bordeaux `#861D2C` on light backgrounds only.** Never on dark.
- **No off-palette color.** The palette is greyscale plus Sunglow (plus Bordeaux on light). No blue, no green, no red that isn't Bordeaux.
- **Premium, technical, forward-looking.** Restraint over decoration, precision over whimsy.

Full palette and type rules: `references/style-dna.md`.

## Workflow

### 1. Read the source

Read the article, doc, deck, Notion page, or brief. Extract:

- The central claim
- Which passages carry a shift in understanding
- What is genuinely structural (and so worth an image)
- What is prose-only and needs no image

Do not illustrate evenly. Choose cognitive anchors: the core judgement, a before/after, an input→output loop, a divide, a ladder of stages, a failure mode, an ownership transfer.

### 2. Propose a shot list first

If the request is strategic ("what should we illustrate", "plan the visuals"), stop at the shot list. Do not generate. For each image give:

- Placement (which section it follows)
- Subject
- The one idea it carries
- Structure type
- What Ami is doing
- Suggested objects
- Suggested labels (ALL-CAPS, ≤3 words each)
- Mode (dark default, or light with reason)

Default 3–6 images. Short pieces get 1–2. Long pieces rarely need more than 8.

### 3. Generate, one image at a time

If the request is explicit ("generate", "make these", "produce the images"), proceed without waiting for confirmation.

Generate each image with a **separate** call — never composite several concepts into one canvas. Use `mcp__claude_ai_Higgsfield__generate_image` when available; otherwise use whatever image generation tool the session provides. Build the prompt from `references/prompt-template.md`.

Every prompt must carry: 16:9, flat Space Black ground, Stone Grey line art, one Sunglow accent, generous negative space, Ami performing the core action, and the explicit prohibitions (no gradient, no glow, no texture, no 3D render, no stock-vector look, no corporate infographic, no off-palette color, no title text).

### 4. Audit the palette — always

Generated images drift off-brand. Amini forbids off-palette color, so every image is audited before delivery. Paths below are relative to this skill's own directory:

```bash
python3 scripts/palette_check.py <image.png>
python3 scripts/palette_check.py <image.png> --mode light
python3 scripts/palette_check.py <image.png> --snap
```

It reports palette coverage, which brand colors are present, rogue hues, Sunglow share, and whether the ground is correct. Exit code is non-zero on failure.

- **Rogue hue reported** (blue, green, teal, non-Bordeaux red) → regenerate with the prohibition stated harder. Snapping will not fix a wrong-colored element.
- **Drift only** (right colors, imprecise hex) → `--snap` writes `<name>-snapped.png` with near-palette pixels pinned to exact brand hex, leaving anti-aliased edges alone.
- **Wrong ground** → the model ignored the dark-first instruction; regenerate.
- **Sunglow over ~15%** → the accent has become a fill; regenerate with fewer accented elements.

Then run the visual checks in `references/qa-checklist.md`.

### 5. Deliver

Save into the working repo as:

```text
assets/<slug>-illustrations/01-subject.png
```

Keep the audited (or snapped) file as the deliverable and preserve originals. Never overwrite existing assets unless asked.

Report: how many images, what each is for, where they are saved, palette audit result per image, and which are strongest versus optional.

## Output register

Keep strategy output short and specific. Keep delivery output factual. Do not explain the style at length — the images carry it.