Form factors for your new AI coworkers — Craig Wattrus, Flatfile

Key Concepts:

• Generative UI
• LLMs (Large Language Models)
• AI Agents
• Flat Files AI Stack
• Invisible, Ambient, Inline, and Conversational AI
• Character Coaching for AI
• Feeling the Material
• Finding the Grain
• Emergence
• Eyes on the Future

1. Introduction: The Convergence of Roles and the Elimination of Traditional Design Processes

The speaker begins by noting the advancements in generative UI tools like V0 and code generation tools like Claude Code. The speaker expresses excitement about the convergence of designers, product people, and engineers working together, envisioning a future where traditional mock-ups and click-through prototypes are replaced by a more hands-on approach. The speaker advocates for "jumping in" and "feeling the material" to discover what emerges through direct interaction with the technology.

2. Flat Files AI Stack Overview

The speaker provides a quick overview of the Flat Files AI stack, which is not an official diagram but represents their perspective. The stack consists of four main buckets:

• Customers' Flat File Applications: Deployed to Flat File's infrastructure.
• Real-Time Context: Data and validation outcomes, including errors and warnings.
• AI Agents: The tools and jobs that AI agents can run.
• User Interface: What is shown to users.

The speaker also categorizes AI interactions into four types:

• Invisible: AI working in the background without direct user interaction (e.g., automatically setting up a demo based on the user's email domain).
• Ambient: AI happening in the space, but not directly worked with (e.g., AI analyzing data in the background and highlighting opportunities for fixes).
• Inline: AI used directly within the user's workflow (e.g., using AI to fix data directly in the data set).
• Conversational: AI interactions through conversation (e.g., using a no-code/low-code agentic system to build flat file apps).

3. Examples of AI Interaction Types

• Invisible AI: When a user signs up for Flat File, AI agents automatically write a flat file application tailored to the user's use case (e.g., an HR demo for a user from an HR company).
• Ambient AI: A tool that analyzes data in the background and highlights opportunities to fix it, indicated by sparkles on the columns.
• Inline AI: Using AI directly within the data to fix it. The AI agents write code that is run on the data set.
• Conversational AI: Using a no-code/low-code agentic system to build flat file apps.

4. Character Coaching for AI Agents

The speaker shares a realization inspired by Amanda Ascal from Anthropic's discussion with Lex Fridman about building Claude's character. The speaker realized the need to shift from controlling AI agents to "character coaching," focusing on building out the desired nature and characteristics of the agents. The speaker created a V0 tool called "chat tuner" to modify the system prompt for their AI orchestrator in build mode, experimenting with different personalities (e.g., more friendly, more balanced, more concise).

5. Theme 1: Feeling the Material

The speaker uses a woodworking analogy to emphasize the importance of understanding the properties of the material (LLMs) before designing with it. The speaker advocates for directly interacting with the models to understand how they work, rather than relying on mockups and prototypes. The speaker's new north star is creating an environment for LLMs to shine, focusing on form factors that help them nail their assignments, stay aligned, and grow as the models improve. The speaker compares LLMs to interns with PhDs, emphasizing the need to provide them with a good "box" (environment) to work in.

The speaker experimented with giving an AI agent a "cursor" in a design tool (Cursor), creating a canvas where the AI could be given orders. While initially enthusiastic, the speaker quickly realized that this approach felt like putting a "Formula 1 driver in a Prius," constraining the AI's capabilities.

6. Theme 2: Finding the Grain

The speaker continues with the woodworking analogy, explaining that "finding the grain" is about feeling out the characteristics of the material (LLM) and the specific piece being worked with. This involves understanding where it is smooth and rough, weak and strong. The speaker emphasizes the need for humility, as the technology is rapidly changing and whatever is built will likely need to be rebuilt.

The speaker shares an example of using the build mode agent to enable the automat plugin, which automatically maps data from the source to the target. While the agent successfully wrote the code, the output was a "wall of text." This led to a redesign of the tool UX, focusing on clear communication, visual cues, and accountability. The redesigned UX includes features like visual indicators of what the agent is doing, confirmation steps, and the ability to express itself (e.g., shaking its head when something goes wrong). The speaker notes that this improved UX can also be applied in other places, such as inline transform functionality.

7. Theme 3: Eyes on the Future

The speaker emphasizes the importance of looking ahead and considering the future of AI. The speaker uses the analogy of a "pelican on a bicycle" to represent a personal area of interest or experimentation. The speaker's "pelican on a bicycle" is autocomplete backed by an LLM. The speaker is experimenting with using LLMs to generate autocomplete suggestions for fixing data, creating a benchmark to evaluate the speed and accuracy of different models. The speaker advocates for designing into the future by building applications that can test desired form factors.

8. Emergence and Unexpected Outcomes

The speaker highlights the importance of "emergence" – the interesting things that can be done with AI that were not possible before. The speaker shares examples of playing with AI and discovering unexpected outcomes:

• An agent combining two files (JSON and CSV) and writing a report without being explicitly asked to do so.
• An agent suggesting that the user contact HR to generate missing employee IDs, rather than attempting to fix the data itself.

9. Conclusion

The speaker concludes by expressing excitement about the new form factors that will be built with AI tools. The speaker encourages the audience to experiment, play, and be curious to discover the full potential of AI.