Geoffrey Litt

Geoffrey Litt works on malleable software: computing environments that people can adapt to their own needs. He works at Notion and explores AI-assisted programming through prototypes and public writing. His central argument is that AI should help people participate in creating software—including understanding and changing it—rather than simply delegate the process to agents.[3][6]

In his July 2026 talk writeup, “Understanding is the new bottleneck”, Litt argues that understanding agent-written code matters beyond checking whether it is correct: “we can understand to participate.” A project involves repeated iterations, and knowledge of the system helps the human conceive its next change. His objection to the software-factory metaphor was similarly explicit in a June 30, 2026 post reproduced by Latent Space: ““Factories” is a depressing vision of the future, metaphors matter”.[6][2]

Litt created /explain-diff to turn code changes into structured explanations in HTML, Markdown or Notion documents. It teaches background concepts, builds intuition before presenting code, and arranges snippets within explanatory prose—a “literate diff”—instead of relying on alphabetical file order. Interactive figures can make concepts tangible; one example lets the reader drag objects around a game garden to understand isometric coordinates. Litt reports using the skill daily and reading its explanation before the raw diff.[6]

He adds a five-question interactive quiz to check whether he has understood a change, drawing inspiration from Andy Matuschak and Michael Nielsen’s work on quizzes embedded in essays. His personal rule is: “I won't send code to others until I can pass the quiz, and I do the same when reviewing others' code.” He describes the quiz as a speed regulator that keeps AI-assisted work from outrunning human understanding.[6]

Another technique is the micro-world, inspired by Seymour Papert: an interactive environment for learning how a system behaves. Litt describes working with an agent to build a debugger for his Prolog interpreter, letting him scrub through execution, inspect the stack and see which rules were evaluated. During a separate website migration, he found Claude’s script difficult to review because he did not know the new framework. He asked Claude to build a command center instead, with buttons to perform the migration step by step and the old and new sites displayed side by side. These tools made the human’s inspection and participation part of the workflow.[6]

Litt also advocates shared spaces where teams can develop a common mental model. He describes running Claude and Cursor agents in Notion and discussing their technical plans with colleagues through comments on collaborative pages. His emphasis is on understanding together, so that people can reason about and develop the work beyond an individual conversation with an agent.[6]

These 2026 practices followed earlier work on giving users more control over software. Before Notion, Litt was a senior researcher at Ink & Switch; he also completed a PhD in human-computer interaction at MIT, advised by Daniel Jackson, and earlier worked in startup design and engineering.[3]

Litt collaborated with Max Schoening, Paul Shen and Paul Sonnentag on Potluck, documented in October 2022. The research prototype explored “gradual enrichment”: starting with a useful freeform note and adding structure and computation as needed, rather than moving the information into a prefabricated app with fixed fields and features.[4]

Potluck’s loop has three parts: extensible searches extract structured information from text; live JavaScript formulas compute with it; and dynamic annotations display results or interactive widgets inside the original document. In the coffee-recipe example, a search finds quantities, formulas multiply them by a scale factor, and a slider controls servings. A pattern such as {number:amount} g finds amounts in grams while exposing the number for computation. The user builds and modifies the tool within the document environment.[4]

The team kept annotations in a separate overlay so the source text remained editable and computed output did not feed back into searches. Explicitly clicked buttons could change the text—for example, recording today’s date after watering a plant. They abandoned manual entity highlighting because it was tedious and introduced hidden state outside the text. JavaScript was a convenience choice, and the prototype expected basic JavaScript knowledge; the team reported informal testing with about a dozen people, mostly programmers. Some widgets, including the default timer, still held ephemeral state outside the text.[4]

The October 2022 essay proposed language models as future assistance for extracting information and writing searches or formulas, while insisting that the process remain “repeatable, inspectable, and understandable.” A later AI | UX meetup account published in April 2023 documents Litt demonstrating AI in Potluck: natural-language spreadsheet editing with a fallback to direct manipulation.[4][1]

In “Malleable software in the age of LLMs”, published in March 2023, Litt argued that LLMs could change who creates software and for what purpose, beyond making professional developers faster. He envisioned ordinary users creating small tools and modifying existing applications. His earlier Wildcard system illustrated both the opportunity and the barriers: users could customize websites through a spreadsheet, but still needed formulas and site-specific scraping adapters. He proposed language-model assistance for those programming tasks while preserving the spreadsheet’s visible, editable interface.[5]

His proposed interaction model pairs an inner loop of direct manipulation with an outer loop of AI-assisted tool modification. Users operate a responsive interface, then ask an LLM for changes when they reach its limits. As he put it, “Chat will never feel like driving a car, no matter how good the bot is.” AI could serve as a local technical collaborator, helping users learn formulas and inspect results. Litt illustrated the idea with a video mockup of an interactive spreadsheet answer, using GPT-generated data that he manually copied into Excel—not a fully automatic spreadsheet product.[5]