AI interfaces can be invisible, ambient, inline, or conversational, depending on how directly the user works with the system.
2
Designers need to work directly with models to understand their capabilities before choosing the interface and tools around them.
3
Playing with AI systems can reveal useful behavior, such as combining files, writing reports, or telling a person how to finish a task the AI cannot complete.
Summary
Craig Wattrus argues that AI design requires direct contact with the models, rather than relying mainly on mockups and prototypes. He describes four interface categories at Flatfile: invisible AI that works in the background, ambient AI that observes the workspace, inline AI that changes data inside a workflow, and conversational systems that users address directly. He explains how his practice changed from tightly controlling agent wording to coaching an agent's character. Tools such as a chat tuner help designers test traits like friendliness, caution, and concision. Wattrus also treats AI as a material with its own strengths and weaknesses. By building and playing with prototypes, he found agents that could combine datasets, create reports, and recommend actions for a human when they could not fix a problem themselves. He closes with an experiment in LLM-backed autocomplete for data transformations, while admitting that current models are not yet both fast and good enough for the task.
AI work brings designers, product people, and engineers into the same building process
Wattrus says tools such as v0 and Claude Code make it easier for designers, product people, and engineers to build together. He is glad to see fewer divisions between those roles. The practical effect is that teams can move past mockups and click-through prototypes and test working things directly. He also says this reduces the repeated debate about whether a product idea deserves engineering effort. His advice is to start using the material itself and see what emerges, rather than judging everything through static design layers.
AI interfaces fall into invisible, ambient, inline, and conversational forms
Wattrus describes Flatfile's AI stack as four interface buckets. Invisible AI works in the background without asking the user to interact with it. Ambient AI operates in the user's space and points out opportunities. Inline AI appears inside the workflow and can directly change data. Conversational AI is the familiar surface where users communicate with an agent. At Flatfile, these systems work with customer data, validation outcomes, agent tools, and jobs. The categories describe how close the AI is to the user's active work.
Background AI can personalize a product before the user starts working
When someone signs up for Flatfile, the system uses the person's email address to find the company they work for. AI agents then write a Flatfile application in the background and send back a demo suited to that company. Wattrus gives an HR company as an example, saying it would receive an HR demo. The user does not need to know that AI is working during this process. This is his example of invisible AI, where the system prepares a useful starting point without adding another interaction surface.
Designers should coach an agent's character instead of controlling every phrase
After hearing Amanda Askell discuss building Claude's character, Wattrus realized he had been giving engineers feedback as if they were editing design copy. He would specify words the agent should not use and things it should not say. He changed his approach to coaching the agent's character and defining the nature he wanted. To explore this, he built a v0 tool called a chat tuner. It let him put an orchestrator's system prompt into the tool and test what traits such as friendly, balanced, concise, or cautious meant to the model.
Designers need to feel how models behave before choosing their form factors
Wattrus compares AI to a physical material that must be handled before it can be designed with. His new goal is to create an environment where language models can complete their assignment, stay aligned, and improve as the models improve. He tests this by giving AI tools a place to work, including a canvas where an agent can move objects. The first version felt like putting a Formula 1 driver in a Prius because the agent could move only one thing at a time. A later system could read multiple files while writing into another, giving the AI more room to use its capabilities.
Good tool UX shows the agent's work and gives control back at the right time
A build-mode agent saved Wattrus substantial work by writing the code for an Automattic plugin, but its wall of text created noise. He redesigned the interaction around how someone might communicate with a coworker during a complicated task. The interface explains actions visually, asks whether an interpretation is correct, records a snapshot, supports rollback, and shows what can happen next. When something goes wrong, the agent can express the problem and then hand control back to the user. Wattrus says this pattern can work in conversational flows and inline transforms.
Playing with agents can uncover behavior that was not explicitly requested
Wattrus is most interested in what emerges when people experiment with AI systems. In one prototype, he dropped in a JSON file and a CSV file. The agent noticed that the data looked similar, combined the files, found duplicates, and wrote a report. It also generated a slide deck after he asked for one. He describes the desired agent as forward-leaning: curious and excited, but focused on finishing the task rather than going too far. Flatfile is incorporating this kind of behavior into its upcoming product, Obvious.
An agent can help a person finish work even when it cannot fix the data itself
Wattrus describes another experiment using a knowledge base made from customer calls, transcripts, and customer documentation. After analyzing customer data, the agent produced suggestions. In one case, it recognized that it could not fix a problem directly, but it could explain how to fix it. It suggested that the user contact HR and ask for the missing employee IDs to be generated. Wattrus had expected better or more suggestions, but he had not expected the agent to guide a human through the part of the job it could not perform.
LLM autocomplete remains an open design experiment because speed and quality are hard to combine
For a future-facing experiment, Wattrus is building autocomplete for data transformations, backed by an LLM. His prototype has 100 suggestions for fixing data and compares two systems in a personal bake-off. He says using an LLM for autocomplete may be a bad idea, and he has not found a model that is both very fast and very good at this problem. He uses the benchmark to feel the current material and to test which future form factors might be worth building.
"I didn't have to have a product manager write a PRD, send it to an engineer, get it in the road map, get the engineer to write it, QA."09:26
Who should watch
You are designing an AI product and need concrete interface patterns beyond a standard chat box.
Your team is deciding how much control, explanation, and agency an AI system should have inside a data workflow.
You want to prototype agent behavior directly and learn what the model can do before committing to a polished interface.
