Design engineering with Maggie Appleton: what engineers can learn from designers in the age of AI agents

The article covers an episode of Gergely Orosz's Pragmatic Engineer Podcast devoted to Maggie Appleton, a staff research engineer at GitHub Next, where she builds prototypes to explore new forms of collaboration between software engineers and AI systems. Appleton places herself at the intersection of design, anthropology and web development, and has been a designer at AI startups such as Elicit (where she was the first designer hired) and lead designer at Normally. The format of the conversation is more visual than usual, as Appleton shows pages from her sketchbook and prototypes during the talk.

One curious biographical detail that appears in the article is that, after graduating in cultural anthropology, one of the career paths open to her was working on the design of torture techniques for the United States military. Appleton turned down that option and moved toward technology. As she explains, her anthropology training has proved useful throughout her technical career, because software is built by people and human relationships matter as much as the code.

Another point in her trajectory is that she trained as a frontend engineer in an unconventional way: for four years she worked as an illustrator at Egghead, a developer education company, rising to art director. To illustrate tutorials she had to genuinely understand what she was drawing —React components, the useEffect hook, JavaScript functions—, which ended up giving her a solid technical grasp. Gergely Orosz mentions that he began following Appleton's work because of her illustrations for Dan Abramov's "Just JavaScript" course, and the article includes an animated example of that work.

One of the most revealing anecdotes of the episode has to do with designing interfaces for AI. In 2021, a year before ChatGPT launched, the startup where Appleton worked invested months of intense work in a "new interface for AI" intended to accelerate scientific research using language models: infinite canvases with cards, Notion-style composable documents, and other interface innovations. The result was that scientific researchers wanted none of it: they preferred the same simple tables they were already familiar and comfortable with. Appleton draws a general lesson from this experience: even when innovating, it pays to start from an interface primitive that is familiar to the user.

The episode also digs into how Appleton understands the relationship between design and software engineering: both disciplines ultimately revolve around problem solving, and the main difference lies in the materials you work with. One aspect she highlights as genuinely difficult is naming: finding the right nouns and verbs to describe new things that people will adopt and use. According to her, this is relatively easy when you are building, say, an online sneaker store, but far more complicated when it comes to a new product for a platform like AWS.

Physical notebooks occupy a central place in Appleton's working method. She usually starts her projects by drawing ideas by hand, because she finds it faster to sketch something on paper than to describe it to a tool like Claude Code. She also points to a practical advantage: an idea drawn in a notebook is still there the next day, whereas an idea fed into an AI tool is much harder to recover after dozens of additional prompts have piled up on top of it.

Another concept of her own that she contributes to the vocabulary of design with AI is "jigs": personal prototypes she builds by asking a coding agent to generate a kind of custom Figma, with controls such as sliders and color pickers that let her adjust parameters in real time —colors, sizes, animation speed—. The term borrows the English word "jig", which in carpentry refers to a device that helps perform a specific task repeatably.

On her current way of working with coding agents, Appleton admits she has stopped reviewing the generated code when a pull request is produced, something she used to do but no longer considers necessary with the current generation of models, at least for building prototypes. Her method consists of, once she is clear on what to build, writing a detailed specification that includes how the agent itself should verify its work.

One of the episode's most quotable reflections has to do with the problems of AI-assisted planning. Appleton describes a frustrating experience in which an agent subjects the user to a multiple-choice interrogation (A, B or C) repeated dozens of times; by around question twenty, fatigue means the brain stops making decisions with any judgment, and since the agent has already suggested which is the recommended option, the user ends up simply accepting that suggestion without thinking. It is, in her words, "a really bad experience" as it stands today.

The article also picks up a term coined by Appleton that has resonated: "capability gaslighting" (roughly, being gaslit about a model's capabilities), which describes the situation in which a frontier model impresses the user into believing it has mastered a task, only to fail spectacularly at that same task the next day. The underlying warning is that we tend to keep trusting models because they previously convinced us of their competence, and that it pays to stay vigilant when working with both models and agents, given how "jagged" their capabilities are.

As for how humans and AI agents should relate going forward, Appleton argues that new kinds of artifacts need to be invented to bridge the two worlds. She describes the agents' universe as one made up of weights, models, skills and protocols such as MCP, as against the human side, made of physicality, texture, light and materials that agents do not understand. Her specific frustration is that an agent cannot "look over her shoulder" while she draws in her notebook, understand what she is sketching and help her push that idea forward.

Finally, the episode offers a practical piece of advice for software engineers: treat the AI agent as a patient tutor for learning design, asking it things like when to adjust line height, what a good sidebar should look like, or how many characters to fit per line. Appleton stresses that, although acquiring product design skills has traditionally been hard for those outside the field, AI agents now make that learning easier.

The article's content closes with the episode's list of references and timestamps (which include sections on the use of Figma, design at GitHub Next, the difference between UX and UI, Appleton's article "One Developer, Two Dozen Agents, Zero Alignment", and her essay "Home-Cooked Software and Barefoot Developers"), as well as links to previous episodes of the podcast on related topics: what "loop engineering" is, design in software engineering with Balint Orosz, whether AI agents really slow work down, vibe coding, and how Codex and Claude Code are built.