monet

Monet — Architecture Image Generation

Generate architectural visualizations using Google Gemini. Each project has a monet/ directory with context, style references, structured prompts (JSON or narrative), and auto-versioned results.

When to Activate

• User invokes /monet
• User asks to generate architectural images or visualizations
• User wants to create image prompts for Gemini

Prerequisites Check

Before any generation, verify:

1. Python 3.7+ available
2. Dependencies installed: google-genai, Pillow, python-dotenv
3. .env file exists in the working directory (or parent) with GOOGLE_API_KEY or GEMINI_API_KEY

If dependencies are missing, run:

pip install google-genai Pillow python-dotenv

If no API key, tell the user:

You need a Gemini API key. Get one at https://aistudio.google.com/apikey Then create a .env file with: GOOGLE_API_KEY=your-key-here

Process

Phase 1: Understand the Project

Check for an existing monet/ directory in the current working directory.

If monet/ exists:

1. Read all files in monet/context/ — this is the project's knowledge base. It contains source documents the user has added (architecture specs, research notes, briefs, technical descriptions, visual identity guides) that inform prompt writing.
2. View images in monet/references/types/ — understand what kind of images this project produces
3. View images in monet/references/styles/ — understand the target aesthetic
4. Check monet/starters/ for base images
5. Summarize: "This project generates [type] images in [style] aesthetic for [purpose]"
6. Proceed to Phase 3 (Author Prompts) or Phase 4 (Generate) depending on user intent

If monet/ does not exist:

• Proceed to Phase 2 (Scaffold)

Phase 2: Scaffold New Project

Create the monet/ directory with standard structure:

mkdir -p monet/{context,references/types,references/styles,starters,prompts,results}

Then help the user set up:

Context (knowledge base):

• Tell the user to add any source documents to monet/context/ — architecture specs, research notes, briefs, technical descriptions, anything Claude should read to write informed prompts
• These files are Claude's reference material, not sent to Gemini

Remind the user to:

• Add source documents to monet/context/ for Claude to read when writing prompts (specs, briefs, visual identity notes, research)
• Drop style reference images into monet/references/styles/
• Drop type reference images into monet/references/types/ (optional)
• Drop starter images into monet/starters/ (optional)

Phase 3: Author Prompts

1. Read all files in monet/context/ — absorb the source material the user has provided (specs, notes, briefs). Use this knowledge to write accurate, informed prompts.
2. Read 1-2 existing prompts in monet/prompts/ (if any) for tone and detail calibration
3. Ask the user what they want to visualize — the image concept
4. If starters exist in monet/starters/, ask if they want to transform one

Drafting the prompt:

Load references/json-schemas.md for template structures.

• JSON prompts (preferred for precise iteration):

  1. Choose the closest JSON template based on image type
  2. Include common fields: scene_type, composition, camera, lighting, constraints
  3. Add type-specific fields — invent fields that fit the image
  4. Show draft to user for field-level review

• Narrative prompts (when user prefers prose):

  • Use photographic/cinematic language (camera angles, lens types, lighting)
  • Include atmosphere cues ("like an architectural magazine editorial")
  • Specify materials and textures precisely
  • Weave constraints into the prose

Save as monet/prompts/<id>.md with this structure:

# Image Prompt — [Title]

## Concept
[What the image represents and why — 2-3 sentences]

## Prompt
[JSON object or narrative text]

Show draft to user for approval before saving.

Phase 4: Generate & Iterate

Generate:

python <skill-path>/scripts/generate.py <cwd>/monet <prompt_id>

Where <skill-path> is the skill's installation directory (resolve from the skill's location).

After generation:

1. Show the generated image to the user using the Read tool
2. Ask for feedback

If user wants changes:

• For JSON prompts: modify specific fields in the prompt file (surgical edits)
• For narrative prompts: rewrite the relevant section
• Regenerate — fresh from clean inputs + updated prompt
• Never use a generated image as input for the next generation (Golden Rule)

Results auto-version: {prompt_id}_v1.png, v2.png, etc. Previous versions are never overwritten.

CLI Script Reference

Script	Purpose	Location
scripts/generate.py	Image generation via Gemini API	Run with Python 3.7+

Usage:

python scripts/generate.py <project-dir> [prompt_ids...] [options]

Options:

• --list — List prompts and version counts
• --model MODEL — Override model (default: gemini-3.1-flash-image-preview)
• --aspect-ratio RATIO — Override (default: 16:9)
• --image-size SIZE — Override (default: 2K)
• --style FILE — Use specific style reference from references/styles/
• --starter FILE — Use specific starter image from starters/

Principles

One Image at a Time

Never generate all prompts at once. Generate one, show to user, get feedback, iterate. Each image deserves attention.

JSON is the Source of Truth

All changes accumulate in the JSON prompt. Every generation is fresh from clean inputs. Never chain through generated images.

Ask Before Assuming

When drafting prompts, show the draft and ask for approval. When the user says "change the lighting," confirm which fields before editing.

Use Context Documents

Read the user's source documents in context/ before writing prompts. The better you understand the subject, the better the prompt.

References

• See references/workflow.md for iteration methodology, conventions, and CLI details
• See references/json-schemas.md for JSON prompt templates (5 scene types)
• See examples/prompt-examples.md for JSON and narrative prompt examples

Boundaries

This skill should NOT:

• Generate all prompts in batch without user review
• Use a previously generated image as input for the next generation
• Modify style/type reference images
• Skip showing generated results to the user
• Write prompts without reading context documents first
• Create prompts without user approval of the draft