Custom GPT System
After taking a prompt-engineering course, I built a frontend for the OpenAI API and started making assistants for different tasks. They share an interface and engine, but each has its own instructions, capabilities, and voice. The project began in winter 2023 and grew through five versions.
The engine handles conversations, image analysis, search, speech, Markdown, and saved chats. Saving a conversation keeps its images, selected bot, and settings together, so I can reopen it on another device.
Making assistants
Each bot is a JavaScript configuration file. Connie helps write those files: I can ask for a calm haiku-writing assistant, review the generated instructions and settings, save the file, and add its ID to the validation list.
The configuration sets the role, output constraints, profile image, temperature, and optional features. Changing one assistant leaves the others alone. The shared frontend uses vanilla JavaScript and CSS variables; PHP handles validation and saved conversations.
What I use them for
At Show of Hands, Dee, Tedium, Amelia, and Gus prepare retail copy, HTML, search descriptions, and tags. That workflow went from about 30 minutes of preparation to roughly 30 seconds in my use.
For music, Gen works with Clouds of Acid, Polite Riot focuses on protest themes, and Jewel Pod helps with melody and structure. Halima reviews writing and emails for grammar and clarity while keeping my voice. Those different uses all run on the same foundation.
Saved chats preserve the context instead of making me start again with each assistant. Conversation persistence and cross-device access were among the additions that came from using the system regularly.
Themes
I grew up with WinAmp and MySpace, where changing the interface was part of the experience. The themes carry some of that into this tool: D0S resembles a CRT terminal, Ugly uses Comic Sans and clashing colors, and Redacted obscures parts of the text. The D0S view below is my favorite.