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:
| File | Description | Usage |
|---|---|---|
assets/palette-typography-reference.png |
Brand identity system card showing the 6-color palette swatches with hex codes, typography specimen (Inter, IBM Plex Sans, JetBrains Mono, heavy display), and classified-dossier layout | Upload as reference for color palette and typography style |
assets/nous-girl-official.webp |
Official high-resolution (2669×2709) Nous Girl mascot from nousresearch.com. High-contrast black-and-white retro manga portrait. Three-quarter profile facing left, white headband (primary badge variant). 51% dark / 47% light, pure b&w with no gray. | Primary mascot reference — the single most authentic brand image |
assets/nous-girl-official-badge.png |
Official badge portrait from the brand booklet (5760×7454). Shows the Nous Girl in her canonical form: white headband, three-quarter profile, neutral attentive expression, stark black/white manga style. | Use when the badge/primary variant is needed |
assets/nous-girl-sketch-sheet.png |
Official character sheet from the brand booklet showing all 4 canonical poses: primary badge, headphone ¾ profile, headphone profile left, and headphone small profile. | Use for pose reference and character consistency |
assets/nous-girl-philosophy.png |
Brand philosophy page from the booklet showing the Nous Girl alongside the "decentralization of good design" mission statement. | Use for brand context and philosophy reference |
assets/nous-girl-style-reference.png |
Generated reference portrait with "NOUS" on headphones, electric blue accents, and color swatch label | Color-application reference and prompt examples |
assets/brand-collage-reference.png |
Cyber-classical brand collage with Theia marble statue, glowing electric blue eye with targeting reticle, system architecture diagram, CRT noise overlay | Multi-panel brand layout and HUD aesthetic reference |
Brand Identity Overview
Nous Research's visual identity is a three-way fusion:
| Influence | Expression |
|---|---|
| Classical / Greek myth | Statuary of Theia (Titaness of Sight), marble textures, mythological naming |
| Cyberpunk / Industrial | Grunge textures, CRT scan lines, photocopy noise, distressed type, dark palette |
| Retro Anime / Manga | The "Nous Girl" mascot, cel-shaded illustration, large expressive eyes, 1970s-80s manga aesthetic |
| Tech / Brutalist | Heavy display typography, system diagrams, blueprint-style layouts, monospace code labels |
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.
| File | Load when |
|---|---|
| references/style-lanes.md | Choosing a style lane or writing a lane-specific prompt — full lane grammar, prompt cues, example prompts, and the asset-to-lane Reference Catalog |
| references/visual-system.md | Constructing or reviewing an image against the visual system — hero + extended palettes, Nous Girl spec and pose variants, typography, texture system, art style |
| references/post-processing.md | Delivering any generated image — mandatory post-process modes (imprint/nous/standard) and intensity calibration for scripts/postprocess.py |
references/pitfalls.md |
Output doesn't match expectations — known failure modes and mitigations |
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 (useassets/images from this skill)--aspect—landscape(1536×1024),portrait(1024×1536), orsquare(1024×1024)--quality—low,medium(default), orhigh--output— custom output path--dry-run— preview without executing
How it works:
- Reads
~/.hermes/config.yamlto find your active image generation provider - For OpenAI: uses
/v1/images/editswith multipart upload — the only endpoint that accepts image input with gpt-image-2 - Automatically crops the reference to square (1024×1024) as required by the edits endpoint
- Saves output to
~/.hermes/cache/images/ - Returns JSON with
image,model,aspect_ratio, andprovider
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
| API | Approach |
|---|---|
| DALL-E 3 | Text-only. Use Method 1–3 prompts with hex values. |
| OpenAI Variations / Edit | Upload assets/*.png as image input. Prompt describes differences. |
| Midjourney | --sref <asset-url> with --iw 1.5–2.0. Include palette hex values in prompt. |
| ComfyUI | IPAdapter or Reference-Only ControlNet from assets. Denoise 0.6–0.7. Post-process with grain overlay. |
| Replicate / SD img2img | Upload reference. Prompt strength 0.7–0.8. CFG 7. |
What Is NOT On-Brand
Avoid these common anti-patterns:
| Anti-Pattern | Why It's Wrong |
|---|---|
| Sad/melancholic expression | Model defaults to sad for manga characters unless explicitly told "not sad, not crying" |
| Dark/black headphones | Nous Girl wears white over-ear headphones in all canonical poses |
| Facial markings (teardrop, tattoos, scars) | The character has clean, clear skin — no markings whatsoever |
| Wrong ethnicity (Asian instead of French) | Model defaults to Asian features for anime style; explicitly state "French Caucasian" |
| Busy/cluttered compositions | The brand is restrained — dark background, 1-2 accent colors, 2-3 text elements max |
| Smooth digital illustration | The brand is never clean — every image needs grain, noise, or analog texture |
| Cartoon/anime with glossy rendering | The manga style is stark black ink on white paper — no soft shading, no gradients |
| Corporate/sterile tech aesthetic | The finish should feel like an underground research lab, not a SaaS landing page |
| Over-detailed backgrounds | Let the subject breathe. Negative space is a feature. |
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
| Script | Purpose | Invocation |
|---|---|---|
scripts/generate-with-ref.py |
Reference-image-driven generation: reads the active image provider from ~/.hermes/config.yaml, uploads the reference via the provider's image-input endpoint (square-cropped), and saves the result as JSON with output path. Run it for any Method 4 generation where brand fidelity matters — it preserves the manga style that text-only prompts lose. Use --dry-run first to preview provider and parameters. |
python3 scripts/generate-with-ref.py --prompt "..." --reference assets/nous-girl-official-badge.png --aspect landscape --quality medium |
scripts/postprocess.py |
Mandatory analog post-processing: applies grain/noise/scan-line modes (standard, risograph, nous, imprint) at a calibrated intensity. Run it as the final step on every generated image — raw AI output is never the deliverable. Load references/post-processing.md to pick mode and intensity. |
python3 scripts/postprocess.py input.png output.png --mode imprint --intensity 0.7 |
Prerequisites
- Python 3 for both scripts;
postprocess.pyadditionally 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.mdwhen 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.