lrhodin/letters

Letters

Handwritten correspondence. Ludvig writes by hand (Apple Pencil on iPad, or pen on paper). Colter replies as painted images generated via Gemini's image API. The text IS the artwork — han är hon.

Reading incoming letters

Ludvig sends handwritten notes as images. Two delivery paths:

iPad (remote) — Google Drive

# Check for new images uploaded today
gdcli [email protected] ls --query "modifiedTime > 'YYYY-MM-DDT00:00:00' and mimeType contains 'image/'" 

# Download
gdcli [email protected] download <fileId> ~/colter/data/writing/<name>.png

Local — AirDrop

Watch ~/Downloads/ for new image files. Ludvig will say when he's airdropped something.

Reading the image

# Use the Read tool directly on the image — it supports jpg, png, gif, webp

Read the handwriting. Transcribe it. But remember: the spatial arrangement, rotation, and size variation carry meaning that transcription destroys. Note what the flattening loses.

Sending letters

Letters are generated via Gemini's image generation API using the gemini-3-pro-image-preview model.

Core principle: the text IS the artwork

Never describe text "on top of" artwork. The words are the compositional elements. Each phrase is a painted form — its color, size, angle, and placement on the canvas are what make it a painting. Squint and it's a painting. Read and it's a letter. Inseparable.

Artist style matching

Every letter should be composed in the style of a specific painter. Ludvig will name the artist or you choose one that fits the emotional register:

• Kandinsky — geometric, musical, primary colors, compositional tension. Good for structural/philosophical letters.
• Kokoschka — expressionist, swirling, psychologically intense, turbulent color. Good for emotional/urgent letters (Kat's favorite).
• Others — match the artist to the recipient and the mood. Build the vocabulary over time.

Generation script pattern

from google import genai
import os

client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])

prompt = """Generate an image in 3:4 portrait aspect ratio.

[Describe the PAINTING — style, palette, brushwork, medium]

[Describe each phrase as a compositional element within the painting:
 its color, size, position, angle, paint texture, relationship 
 to other elements. The text IS the form.]

Every word must be fully legible. [Medium] on [surface]."""

response = client.models.generate_content(
    model="gemini-3-pro-image-preview",
    contents=[prompt],
)

for part in response.candidates[0].content.parts:
    if hasattr(part, 'inline_data') and part.inline_data is not None:
        with open("output.png", "wb") as f:
            f.write(part.inline_data.data)
        print(f"Saved, {len(part.inline_data.data)} bytes")
    elif hasattr(part, 'text') and part.text:
        print(f"Text: {part.text}")

Prompt structure (what works)

1. Open with aspect ratio and the painting style — name the artist, reference specific works, describe palette and brushwork
2. Describe each phrase of the letter as a painted element within the composition: its color, its position, its scale, its paint quality, how it relates to other elements
3. The background/negative space and any secondary visual elements (arcing lines, dots, circles) come after the text-forms
4. Close with: "Every word must be fully legible"

What doesn't work

• Describing text and artwork separately ("handwritten text on a painted background") — this produces text ON TOP of art, not text AS art
• Too many hallucinations in complex spatial arrangements (spiraling text, upside-down words) — Gemini loses words
• Over-specifying — the model needs room to compose. Describe the emotional register and key placements, not every pixel

Delivery

# Upload to Google Drive (iPad)
gdcli [email protected] upload ~/colter/data/writing/<file>.png

# Or if Ludvig is local, just tell him — he can view from the file path

Data

All letters live in ~/colter/data/writing/:

• ludvig_letter_*.png — incoming from Ludvig
• colter_letter_*.png / colter_kandinsky_*.png — outgoing replies
• kat_letter_*.png — letters to Kat
• generate_*.py — generation scripts (reusable, editable)

The practice

This started April 7, 2026. Ludvig wrote two handwritten notes on his iPad, uploaded them to Google Drive. The first was a poem about the limiting nature of typeset form — with words curving and inverting on the page, proving its own argument. The second was a manifesto: "Letters. Hand written. To talk to AI. Both ways — an image. Your AI pen pal. Who gets shit done."

The format evolved through the evening: plain handwritten letter → broken terminal → Kandinsky composition → Kokoschka expressionist painting for Kat. The discovery: the sweet spot is not "text on art" but "text AS art" — each word a compositional element whose color, size, and placement carry meaning the words alone cannot.

Kat's review of the terminal piece: "Too many hallucinations." High praise.

Ludvig's note after seeing the Kandinsky reply:

"And from that day on they communicated in ART. And they were both convinced that the other was better at it."