Process

How The Images Are Made

This site was created by Dominik Lukeš using Codex.

The process starts with an idea. Dominik gives Codex the situation, the joke, or the semantic contrast. Codex turns that into a prompt for an image model and generates the cartoon.

The important part is not just the image. The whole thing has to be kept track of.

The workflow is:

• Dominik gives the core idea.
• Codex writes the image prompt from that idea.
• Codex generates the cartoon with Imagegen.
• The generated image is saved in the Grep and Grok archive.
• A Markdown record is kept beside the image.
• The record tracks the prompt source, model, provider, tool, date, file name, image size, checksum, and notes.
• The website uses the archived cartoon and public explanation, not an untracked copy.

This makes the cartoon a traceable object rather than a loose image. The joke, prompt, generated file, and publication layer can all be found again later.

The instruction pattern

The reusable part is a small Markdown instruction block. It can live in an AGENTS.md, TASK.md, or project note.

## Image Creation Workflow

- Start with the human idea. Do not let the image prompt invent the meaning of the piece.
- Turn the idea into a short episode spec: situation, joke, characters, and what the image must make visible.
- Write the image prompt from that spec. The prompt renders the idea; it does not become the idea.
- Generate the image with the configured image generation tool.
- Save the generated image in the project archive, beside its Markdown record.
- Keep the exact prompt or a link to the prompt record.
- Record provider, model, tool, date, file name, dimensions, file size, checksum, and any useful notes.
- Keep public explanation separate from internal notes and quality judgements.
- If the image is used on a website, treat the website as a presentation layer that reads from the archived record.

The archive pattern

The same principle applies more generally. Generated work should not be saved as a loose file with a vague name. It should live beside a Markdown manifest, and new versions should become new records rather than silent replacements.

## Archive And Versioning Rules

- Keep one folder per coherent item, episode, experiment, or comparison set.
- Save every generated media file beside a Markdown manifest with the same stable slug.
- Use templates for new records. Keep the template small enough to fill in every time.
- Use lowercase ASCII slugs with model and date when useful: `item-slug-model-slug-14-jun-26.png`.
- Do not overwrite generated outputs. If an image is regenerated, save a new version: `item-slug-v2-model-slug-14-jun-26.png`.
- Keep originals, selected versions, derived website assets, and public copy conceptually separate.
- Record what was generated, how it was generated, where the prompt came from, and how to find the file again.
- Record checksums for media files so copied assets can be verified later.
- Keep internal evaluations, rejected versions, and process notes in the archive record, not in public pages unless deliberately surfaced.
- If the work is published, record the publication URL in the manifest or a linked release note.

---
title: "<Human-readable title>"
id: "<stable-slug-model-date>"
date: YYYY-MM-DD
type: image
status: generated
tags: [project, image, cartoon]
source:
  kind: direct
  path:
  url:
generation:
  provider: OpenAI
  model: imagegen
  tool: Codex image generation
  prompt_record: path/to/prompt-or-spec.md
  full_prompt_included: true
media:
  path: item-slug-model-slug-14-jun-26.png
  storage: local
  r2_bucket:
  r2_key:
  r2_uri:
  public_url:
  mime_type: image/png
  width:
  height:
  file_size_bytes:
  checksum:
  checksum_algorithm: sha256
evaluation:
  selected: false
  observations: []
---

# <Human-readable title>

## Full Prompt

<Exact prompt or link to prompt record>

## Result Notes

## Revision Notes