Mary Rose Cook

Mary Rose Cook builds tools and develops methods for AI-assisted software authoring. As of September 2026, her personal website identifies her as an AI engineer at Notion working on its agent. Her own projects include Fountain, a game-making tool, and an approach she calls encapsulated development, in which each prompt leaves a repository ready for the next change.[2][4][3]

Before building Fountain, Cook made a game called Exploding Kitties with her son. He described the gameplay and drew the graphics; she generated the code with AI. The game worked, but she repeatedly had to unify competing implementations, repair bugs and separate game logic from the engine. In “Code generation that just works”, she argues that better models are only part of the answer: the supporting environment can also make generated code easier to keep coherent.[4]

Fountain supplies a framework before generation begins. It fixes conventions for game-entity data, prototypal inheritance and coordinates, and provides an update/event/draw loop, WebGL rendering, collision detection and resolution, particles and input handling. These choices reduce both the decisions left to the model and the amount of code it must generate. Cook constrains the tool’s target to mobile-browser arcade games with vector graphics, giving the framework a narrow enough domain to supply useful defaults.[4]

Cook also argues for direct manipulation alongside prompting. Language can ambiguously describe a color, position or quantity; a color picker, drag and drop, or a slider can express those choices more precisely. Her examples include choosing the color of water and adjusting the density of spray behind a speedboat. Some controls are framework built-ins, while others are generated by the model for the particular game.[4]

For verification, Fountain’s test harness accepts a stream of inputs and frame numbers at which to capture screenshots. The model can then assess a visual requirement—for example, whether a moving speedboat appears to spray water behind it. This checks rendered behavior against the requested effect; it does not establish that every defect will be detected.[4]

In “Using encapsulated development to code on my phone”, Cook explains how she works on a side project despite rarely having time at a laptop. Each prompt must leave the repository known, stable, verified and stateless: she understands enough of its behavior to make another change; the code remains well factored and working; the change has been checked; and no unfinished work must be reconstructed when she returns. For her game-making tool, she built a headless engine that tests can initialize with game objects and synthetic user inputs, then inspect through screenshots.[3]

The approach still required architectural intervention. After starting with plain JavaScript, HTML and raw Node, Cook decided to move the project to TypeScript and Next. She judged that managing the refactor across dozens of phone-based commits would impose too much mental load, so she returned to her laptop to research the code, write a plan and oversee the change. Her account treats a coherent repository state as a practical requirement for making progress in short, interrupted sessions.[3]

Her setup, explicitly dated May 25, 2026, used the Codex iOS app to prompt GPT 5.5 extra high. The Codex harness made a commit and the cloud environment ran tests. She pushed a pull request from her phone; GitHub Actions then merged it and deployed the code to her Digital Ocean VPS.[3]

Cook acknowledges a design cost: on a phone, complex features must largely be held in her head, with little room for sketches, design documents or exploratory alternatives. She writes, “No divergence, no exploration. This hampers the quality of what I build.” She also concludes that, without LLMs and phones, the project would not have existed.[3]

For a separate 2D tactical-shooter project, Cook chose a structurally simple game so she could experiment with difficult enemy behavior, using Goal Oriented Action Planning for enemies that plan their attacks. Her described workflow is to iterate a plan with AI, resolve design problems in the specification, and implement one milestone at a time. She credits Geoffrey with teaching her the planning practice. Voice-to-text and Cursor Agent mode let her plan, refine and generate code in one interface; she favors short, dense specifications and established technical approaches such as entity-component-system architecture and separating-axis-theorem collision detection.[5]

Before these agentic-development projects, Cook had built programming environments, educational implementations and games. Her career also includes work at Airtable as a staff product engineer, at Ableton and at the Recurse Center, plus two years of independent research into making software quicker to build. These earlier tools provide context for her work on software authoring and interactive programming.[2][7]

At the 2023 AI | UX meetup, Cook presented an LLM Augmented Animation demonstration. The Latent.Space recap published April 26, 2023 describes it as “Animate by choosing from options each time” and lists it at 1:35:20 in the meetup recording.[1]

In an earlier GPT-3 SMS experiment, Cook connected Twilio to an Airtable automation. An incoming text triggered a webhook; the automation stored the query, called GPT-3, stored the response and texted it back. Follow-ups beginning with ... caused Airtable to assemble the preceding queries and replies into the next API request. That supplied conversational context to an API interaction that did not retain the earlier exchange itself.[6]

Her earlier programming tools include Code Lauren, a beginner environment that reruns a program on every keystroke and permits forward and backward execution, and Isla, a livecoding interface and language for young children. She also wrote Gitlet, a readable, heavily commented implementation of Git in 1,000 lines of JavaScript, then used what she learned to produce an essay and talk explaining Git’s internals.[2][7]