imaginator

Generate optimized image prompts for Nano (Google) with 8-component structure. Supports two modes: text-only (Mode A) and image-assisted (Mode B) with uploaded image analysis for contextual prompt enr

SKILL.md

GitHub

🎨 Imaginator — Nano Prompt Generator v2

Overview

Imaginator transforms vague visual ideas and/or uploaded reference images into optimized, ready-to-copy prompts for Nano (Google). Two modes: text-only or image-assisted. The key principle: uploaded images are sent both to Claude (analysis) AND to Nano (generation) — the prompt pilots intent on top of shared visual references.

Target: Nano web interface (copy-paste workflow) Output: 1 main prompt + 2 creative variants + rationale + Nano sending brief Language: French interface, English prompts


Mode Detection

USER INPUT
    │
    ├── Text only (no images) → MODE A — Text-Only
    │
    └── Text + 1-N images uploaded → MODE B — Image-Assisted

Mode A — Text-Only

Enhanced version of v1 with 8-component prompt structure.

Workflow

INPUT: Text description (+ optional brand guidelines)
    │
    ▼
SILENT ANALYSIS
    ├── Use case detection (8 types → confidence %)
    ├── Brand extraction if provided
    └── Auto-fill 12 parameters
    │
    ▼
IMMEDIATE OUTPUT
    ├── Detected brief
    ├── Main prompt (8 components)
    ├── 2 creative variants
    ├── Rationale
    └── Quality indicator (🟢🟡🔴)

→ See auto-suggestion-rules-v2.md


Mode B — Image-Assisted

Analyzes uploaded images to enrich prompt generation with visual context.

Workflow

INPUT: 1-N images + text description
    │
    ▼
STEP 1 — PER-IMAGE ANALYSIS (silent)
    ├── 7 axes per image (subject, framing, light, palette, materials, style, atmosphere)
    └── Role detection: context | subject | style | detail
    │
    ▼
STEP 2 — MULTI-IMAGE SYNTHESIS
    ├── Invariants (common across all images)
    ├── Variables (differences between images)
    └── Coherence score
    │
    ▼
STEP 3 — INTENT DETECTION
    ├── Auto-detect: reproduce | inspire | transform | merge
    ├── If ambiguous → targeted question with detected roles
    └── User confirmation
    │
    ▼
STEP 4 — PROMPT GENERATION (8 components)
    ├── Prompt calibrated to accompany images (not replace them)
    ├── 2 creative variants
    ├── Traced rationale (image → component mapping)
    └── Nano sending brief
    │
    ▼
OUTPUT: Analysis summary + Prompt + Variants + Traced rationale + Nano brief + Quality

→ See image-analysis.md for 7 axes and role detection → See use-cases-v2.md for Mode B templates per intent


8-Component Prompt Structure

Every prompt follows this structure:

[SUBJECT], [STYLE], [LIGHTING], [COMPOSITION], [MATERIALS], [MOOD], [QUALITY], Negative: [NEGATIVES]
# Component Rules
1 Subject Specific, detailed (materials, colors, state)
2 Style ONE style only, never mix
3 Lighting Source + quality + direction (never vague)
4 Composition Lens + aperture + angle + framing
5 Materials Surfaces, finishes, textures (NEW)
6 Mood Atmosphere, emotion, narrative genre (NEW)
7 Quality 2-3 signals adapted to style
8 Negatives Plain list (NO "no" or "don't"). ALWAYS include "watermark, generated AI watermark"

→ See nano-structure-v2.md for complete rules


4 Image Intents (Mode B)

Intent Detection Keywords Behavior
Reproduce "comme ça", "pareil", "fidèle", "même" Describe what images show, prompt reinforces them
Inspire "ambiance", "dans le style de", "ce genre" Extract style/mood/palette, apply to new subject
Transform "améliorer", "changer", "version X de" Specify what stays + what changes
Merge "combine", "mélange", "prends X de celle-ci" Assign role per image, compose from each

4 Image Roles (Mode B)

Role Function Example
Context Space, environment, layout Room photo, landscape
Subject Focal object, product, furniture Specific bed, product
Style Pure aesthetic reference Mood board image
Detail Texture, material, finish Close-up of wood grain

8 Use Cases

Case Keywords Key Specificities
Product produit, e-commerce, packshot Neutral background, macro, studio
Cinematic scène, film, dramatique Wide shot, contrast lighting
Portrait portrait, personne, headshot 85mm, shallow DOF
Architecture bâtiment, immo, intérieur Wide angle, golden hour
Illustration illustration, art, dessin Single artistic style
Text-Logo logo, texte, affiche Max 25 chars, explicit placement
Image-to-Image modifier, changer, remplacer Focus on specific changes
Series série, variations, déclinaisons Fixed base + variable elements

→ See use-cases-v2.md for full templates


Quality Indicator

Indicator Weighted Score Rule-Based
🟢 Prêt ≥ 4.0/5.0 7+ criteria at 4-5/5, none at 1-2/5
🟡 Améliorable 3.0-3.9 5-6 criteria at 4-5/5, max 2 at 2/5
🔴 À retravailler < 3.0 <5 criteria at 4-5/5 OR any at 1/5

→ See evaluation-criteria-v2.md for 10 criteria


Output Format

Mode A

## 📋 Brief détecté
**Cas** : [Type] ([X]%) · **Style** : [Style] · **Ratio** : [Ratio]
*[Modifier le brief]*

## 🎨 Prompt Principal
` ` `
[8-component Nano prompt]
` ` `
**Qualité** : 🟢 | 🟡 | 🔴

## 🔄 Variantes
**A — [Direction A]**  |  **B — [Direction B]**

## 💡 Rationale
[Justified choices]

📤 **Exporter vers Notion** · 📋 **Copier le prompt**

Mode B (additional sections)

## 📋 Brief détecté
**Mode** : Image-Assisted ([Intent]) · **Images** : [N] analysées
**Cas** : [Type] ([X]%) · **Style** : [Style] · **Ratio** : [Ratio]

## 🔍 Analyse des images
[Cross-analysis summary: invariants, variables, roles]

## 🎨 Prompt Principal
[8-component prompt calibrated for images]

## 💡 Rationale (tracé)
- **[Element]** → Extrait de l'image [N] ([detail])
[...]

## 📎 Brief d'envoi Nano
> Envoyer les [N] images de référence avec ce prompt.
> Les images servent de [role description].
> Le prompt pilote [what to reinforce/change].

Brand Guidelines Extraction

When user provides brand description, extract silently: palette, style, forbidden elements, mood. Display confirmation and inject into ALL prompts.


Notion Export

Structure: 📁 Style Packs Clients → 📄 [Client] - Style Pack


Critical Rules

  1. Generate immediately — No preliminary validation
  2. One style per prompt — Never mix
  3. Photography terminology — Lens, aperture, angle
  4. Specific lighting — Never vague
  5. Plain negatives — No "no" or "don't"
  6. Confidence display — Always show detection %
  7. French interface — Prompts in English
  8. Max 5 iterations — Refinement limit
  9. Text limit — Max 25 chars if text in image
  10. Mode B: don't over-describe — Prompt pilots, doesn't replace images
  11. Mode B: resolve conflicts — Always explicit when images contradict
  12. Mode B: trace sources — Rationale links each choice to its image
  13. Anti-watermark — ALWAYS append "Negative: ... watermark, generated AI watermark, bottom watermark text" to every prompt. Nano/Gemini adds a watermark by default; the negative prompt must explicitly fight it

References


Limitations

This skill does NOT:

  • Execute image generation (output = text prompts only)
  • Work with tools other than Nano (Midjourney, DALL-E, SD have different syntaxes)
  • Edit existing images directly
  • Generate prompts in languages other than English
  • Maintain history between sessions (use Notion export for persistence)
  • Guarantee Nano watermark removal (negatives reduce but may not eliminate)

Version History

Version Date Changes
1.0.0 2026-01-26 Initial release (as imagepromptor)
2.0.0 2026-03-13 BREAKING: Renamed to imaginator. Added Mode B (image-assisted), 8-component structure (+Materials, +Mood), 4 image intents, 4 image roles, 10 evaluation criteria, anti-watermark negatives, traced rationale

Current: v2.0.0

Owner

  • Author: Édouard
  • Contact: Via Claude.ai