Font Selector ｜ Fabrie Write(AI)

👷editing…

Dashboard screen showing navigation controls in a self-driving vehicle, including a large tablet displaying a map view and vehicle information, with a red header labeled 'Navigation under autopilot/autosteer'.

AI Assistant:
Fabrie Write

Product design: Muki
UX design: Muki
UI design: Mao, Gaoyi
Prompt Engineer: Yiqi, Xiran, Muki

01

Framing

In 2023, after a year of launching the Fabrie whiteboard and completing a major technical refactor, the team started experimenting with AI features. Early tests (v1.0) included basic image-related capabilities such as text-to-image, image-to-image, and super-resolution. These experiments validated user interest, but the features remained lightweight and loosely connected to the core whiteboard experience.

A timeline diagram illustrating the evolution of image generation technology from early tests and text-to-image approaches to consolidation and expansion with Canvas-GPT, image generation, text processing, and image editing.

As Midjourney and ChatGPT gained massive traction, many tools expanded toward “AI + whiteboard” scenarios. For Fabrie, the whiteboard was already a natural canvas for AI-generated content, presenting an opportunity for feature expansion and product differentiation.

A computer screen displaying a chat interface with various menu options in Chinese, and a dark theme with an icon on the left side.

ChatGPT - chat-style, timeline-based interactions

Screenshot of a digital brain storming board titled 'Research on autopilot experience' with various notes, images, and graphs discussing autonomous vehicle features, user awareness of rules, and data projections for 2023.

fabrie - where AI meets whiteboards for spatial collaboration

However, most AI features on the market were limited to chat-style, timeline-based interactions, focusing mainly on text or image generation. Fabrie’s real strength lies in spatial collaboration. Simply embedding a chat window into the whiteboard would neither create differentiation nor bring meaningful value to our users.

Challenge

How might we leverage Fabrie’s spatial strengths to design AI features that truly enhance collaboration and document value—rather than becoming just another replaceable chat plugin?

02

Dicovering

I analyzed several AI tools (ChatGPT, Notion AI, Gamma, Tome, Miro, etc.) to understand differences in feature types, interaction patterns, and integration with core products.

  • Mainstream interactions are concentrated in chat windows or command-based sidebars

  • Most features focus on content generation, with limited integration into spatial objects

Two Core Capabilities of AI

Based on this research, AI capabilities can be broken down into two steps:

Placeholder graphic with a message requesting generation of an image, text, video, or code about a specific topic.

Step 1: Natural Language Processing — understanding user text input

A simple diagram showing a curved arrow pointing downward and to the left.
Screenshot of a chat with generated code snippet in Python that defines a greeting function.

Step 2: Content Generation — producing images, text, low-code, executable code, or video

Create from scratch

Close-up of a black circular object with an arrow pointing downward, set against a plain white background.

Within Step2: Content Generation, we identified two distinct modes:

  • Create from scratch: generating entirely new objects (e.g., sticky notes, mind maps)

  • Enhance existing: modifying current objects (e.g., change color, adjust style)

Interactive: Traditional vs AI

Comparison of two methods for changing the color of a yellow sticky note from yellow to purple through interactive steps, including a color palette and a button labeled "Go" for the new method.

Traditional commands rely on fixed buttons (e.g., pressing “Back” to return to the previous page)

AI commands rely on understanding natural language, breaking the one-to-one mapping between command and task → offering greater flexibility but also higher uncertainty

03

Ideation

Challenge

In v2.0, the challenge became concrete: How might we introduce natural language input as a new interaction model, while making AI capabilities feel native to whiteboard objects and spatial workflows?

Comparison chart showing version 1.0 labeled 'Early tests' with features such as text-to-image, image-to-image, super-resolution, and version 2.0 labeled 'Consolidation & Expansion' with features including Canvas GPT, image generation, text processing, and image editing.
A flowchart comparing feature complexity and interaction needs, from low customization with button trigger to high customization with natural language input, including text processing, image editing, image generation, and template generation.

We approached this challenge from a function-first perspective.

Feature scope (by category):

  • Image Generation: generate images based on prompts

  • Text Processing: summarize, continue writing, extract tasks, expand details, split paragraphs, translate, refine text

  • Image Editing: background removal, OCR text recognition, upscaling, AI-assisted edits

  • Template Generation (Canvas GPT): mind maps, brainstorming, pros & cons, business model canvas, stakeholder maps, timelines, data tables

From these functions, I categorized interaction needs into two types:

  1. Low-customization needs (e.g., translation, background removal) → retain button-triggered actions

  2. High-customization needs (e.g., generating complex illustrations, custom templates) → introduce natural language input