Imported from magnus919/agent-skills (nous-branding/SKILL.md). Install upstream with npx skills add magnus919/agent-skills --skill nous-branding. Copyright stays with the author (MIT).

Nous Branding

Generate images and brand-consistent visual content inspired by Nous Research ("The AI accelerator company").

When not to use

• Non-Nous design work — this skill encodes one specific brand system; do not apply its palette, textures, mascot, or compliance rules to unrelated brands or generic design requests.
• Official brand representation — the Nous Girl mascot is not a substitute for the logo; use the Nous Research wordmark/symbol for official brand use.

Reference Images

This skill ships reference images in assets/ that can be used as visual anchors for img2img, style transfer, image variation, or prompt construction:

Brand Identity Overview

Nous Research's visual identity is a three-way fusion:

Tagline: "The AI accelerator company" Key phrases: "Advance human rights and freedoms", "Open source language models", "Unrestricted availability and use" Vibe: Intellectual but gritty — a cutting-edge research lab operating in the shadows

Loading Guide

Load references on demand — do not load everything at once.

Image Prompt Templates

Method 1: Full Brand Portrait

A cyber-classical brand identity illustration in the style of Nous Research / Project Theia.
[SUBJECT DESCRIPTION]. High-contrast dramatic lighting with deep near-black background (#00000E).
Electric blue (#3847FF) primary accent. Soft lavender (#BDA6FF) and burnt orange (#D6825A)
secondary accents. Deep teal (#2E706B) shadow tones. Overlaid with risograph grain texture,
photocopy noise, and subtle CRT scan lines. Retro anime cel-shading combined with neo-classical
sculptural forms. Geometric HUD overlay lines in burnt orange. Bold, distressed display typography.
Raw, analog, imperfect finish. No corporate polish.
Palette: #00000E bg, #3847FF accent, #BDA6FF secondary, #D6825A warm, #E6E6E6 text.

Method 2: Nous Girl Mascot

High-contrast retro manga anime portrait, 1970s-80s cel-shaded style.
A young woman with large anime eyes, shoulder-length dark hair with blunt
straight-across bangs. White over-ear headphones. Three-quarter profile facing left.
Melancholic introspective expression. Bold heavy outlines. Pure black and white with
no grayscale. [Optional: Electric blue #3847FF hair highlights for color version].

Method 3: Brand System Sheet / Collage

Multi-panel brand identity system sheet in Nous Research / Project Theia style.
Grid layout. [Describe panels]. Color palette: #3847FF electric blue, #BDA6FF lavender,
#D6825A burnt orange, #2E706B deep teal, #E6E6E6 off-white, #00000E near-black.
Texture swatches: risograph grain, photocopy noise, CRT scan lines, paper fiber, ink smudge.
Typography: heavy distressed display for titles, Inter/IBM Plex Sans for labels,
JetBrains Mono for technical data. Grunge textures throughout. Dark near-black background.

Method 4: Reference-Image-Driven Generation (Recommended)

This is the preferred method for generating brand-consistent images. Use the scripts/generate-with-ref.py script which reads your Hermes config, determines the active image provider, and hits the API directly with the reference image as contextual input — bypassing the built-in image_generate tool which only supports text prompts.

python3 scripts/generate-with-ref.py \
  --prompt "Your prompt describing the desired image" \
  --reference assets/nous-girl-official-badge.png \
  --aspect landscape \
  --quality medium

Features:

• --prompt (required) — image description
• --reference (required) — path to a reference image (use assets/ images from this skill)
• --aspect — landscape (1536×1024), portrait (1024×1536), or square (1024×1024)
• --quality — low, medium (default), or high
• --output — custom output path
• --dry-run — preview without executing

How it works:

1. Reads ~/.hermes/config.yaml to find your active image generation provider
2. For OpenAI: uses /v1/images/edits with multipart upload — the only endpoint that accepts image input with gpt-image-2
3. Automatically crops the reference to square (1024×1024) as required by the edits endpoint
4. Saves output to ~/.hermes/cache/images/
5. Returns JSON with image, model, aspect_ratio, and provider

Why this matters: Text-only generation loses the precise manga style, character proportions, and contrast balance of the Nous Girl. Uploading the official badge preserves the specific 1970s–80s cel-shaded ink style.

Prompting for reference workflows: State what to preserve from the reference, then what to add:

• "Keep the character's white over-ear headphones, white collared shirt, solid black hair with blunt bangs, neutral expression"
• "Maintain the same high-contrast black ink on white manga style"
• Then add the scene details, text, lighting, etc.

Prompt Formula

[STYLE: cyber-classical / Nous Research]
+ [SUBJECT DESCRIPTION]
+ [PALETTE: #00000E bg, #3847FF accent, #BDA6FF, #D6825A]
+ [TEXTURES: risograph grain, photocopy noise, CRT scan lines, paper fiber, ink smudge]
+ [LIGHTING: high-contrast chiaroscuro, dramatic spot, neon edge highlights]
+ [MOOD: intellectual, gritty, underground, calm/attentive]
+ [TYPOGRAPHY: heavy distressed display, Inter/IBM Plex Sans labels, JetBrains Mono code]

Post-Processing

Raw AI-generated images are too clean for the Nous aesthetic. Post-processing is mandatory after every generation — the raw generated image is never the final deliverable.

python3 scripts/postprocess.py input.png output.png --mode imprint --intensity 0.7

Run it as the final step after any generation method (text-only, img2img, multi-pass, or any provider). For mode selection (imprint / nous / standard) and intensity calibration per output type, load references/post-processing.md.

API Workflow Notes

What Is NOT On-Brand

Avoid these common anti-patterns:

For a complete list of known failure modes and mitigations, see references/pitfalls.md.

Brand Compliance Checklist

• Background is near-black (#00000E) or very dark
• Electric blue (#3847FF) is used as primary accent
• At least one grunge texture visibly applied (grain, noise, scan lines, paper, ink)
• High contrast — dramatic light/dark difference
• Palette is restricted to the specified colors
• If mascot appears: white headphones, manga style, neutral attentive expression, three-quarter profile
• If text appears: heavy distressed display for titles, clean sans for labels, monospace for code
• No flat/clean/corporate polish — finish is raw and tactile
• Overall impression: intellectual, gritty, underground research lab

Full color tables, the complete Nous Girl mascot specification with pose variants, typography roles, texture definitions, and art-style attributes live in references/visual-system.md; lane-specific prompt construction lives in references/style-lanes.md.

Available Scripts

Prerequisites

• Python 3 for both scripts; postprocess.py additionally needs an imaging backend (Pillow) available.
• For generate-with-ref.py: a configured Hermes image-generation provider in ~/.hermes/config.yaml (or an explicit --provider) whose API accepts image inputs — text-only endpoints cannot do reference-driven generation.
• The bundled reference images in assets/, which serve as img2img anchors; pass one via --reference.
• An agent or client capable of image generation when working outside the scripts (per compatibility).

Limitations

• This skill encodes one specific brand system — the palette, mascot rules, textures, and compliance checklist here are not general design guidance and must not be applied to other brands (see When not to use).
• The Nous Girl is not a logo substitute; official brand representation requires the wordmark/symbol, not generated mascot art.
• Generated imagery approximates the style even with references: always run the Brand Compliance Checklist before delivering, and consult references/pitfalls.md when output drifts (sad expressions, dark headphones, glossy rendering).
• Model/provider behavior changes over time — flag-specific details like endpoint names may drift from what the scripts assume; verify against your configured provider.