Building Idyllic - Prior IterationsThe earlier Idyllic designs, with original sketches and prototypes from 2024 and 2025.

I started building Idyllic in 2024 to make AI tools that people could adapt to their own work. A useful conversation often depended on context I had supplied and a sequence of questions I had worked out. I wanted to save that work in an application I could keep using and change as I learned what it needed.

Julius Lattke worked with me on the product designs. We explored small applications with their own feeds, documents that could run a procedure, and agents that shared editable objects with the person using them. I built prototypes to work through the behavior behind those interfaces.

The articles below cover the 2024 and 2025 iterations. They use the original sketches and interface captures to explain the problem each version addressed and the decisions that shaped it. Captions distinguish proposed designs from prototype behavior.

2024

From Telegram workflows to AI programs begins with an assistant for preparing design proposals. The question was how to connect a model response to the information and steps surrounding it.

Interactive results for AI apps covers the feed-based applications and Julius's Search Agent. A running workflow needed somewhere to ask for input and show the sources behind its answer.

Clarity Bear: From idea to specification follows Clarity Bear. I wanted the conversation to collect missing information and leave me with a document I could work from.

A visual builder for AI apps covers the visual program builder and the experiments with editable objects inside documents. Buttons and generated results needed behavior that the application could represent explicitly.

2025

A task list shared with an agent follows the planner designs. Changes made through conversation needed to affect the same tasks I could edit through the interface.

Documents that run AI procedures covers the document editor and Julius's execution movie. Instructions, references and intermediate results had to remain understandable while the program ran.

Controls for agent workflows follows the builder and mobile interfaces. An agent needed controls suited to its task, including confirmations and editable documents.

A language for AI applications covers the work on representing an application's concepts and operations in a form a model could use. The language and builder had to agree on what those operations meant.

A shared runtime for concurrent agents follows the move to TypeScript classes and the Virtual Office prototype. The browser and the agents operated on state belonging to the same running application.