Does a human still need
to read what the agent wrote

He concedes the pressure behind that expectation. Agents are writing a growing share of the code, and keeping up with their output is becoming genuinely difficult. But he argues that reading every diff line by line is not the only way to understand a system. A diff is the before-and-after view of a code change, typically the object examined during review.

The talk introduces three other ways of building understanding: code explainer docs, quizzes that test his own understanding, and micro-worlds he can explore and manipulate. Those three make up the sections that follow.

A diff is the raw record of which lines changed, and it is not organized for explaining. Litt starts from a different question: what would the best possible explanation look like if a human or AI team had carefully worked through how to explain the change? That question set the direction for the skill. It is his personal skill, published as a gist, rather than an official Notion or Anthropic product; he uses it every day and says many of his coworkers have found it valuable.

The skill can produce explainers in three forms: HTML, markdown, or Notion docs. He describes Notion as a useful place for teams to collaborate on and discuss these documents, while disclosing that he works at Notion and is therefore biased.

He demonstrates the structure with a change to the visual perspective of a video game. The document begins by describing the existing game engine. Before any code, it states the goal as "make the garden feel three-dimensional with 2D drawing tricks" and then explains the concepts behind it, one of which is isometric projection, a way of drawing a two-dimensional scene so that it reads as three-dimensional, commonly used in games. In his account, this sequence catches him up as the human, giving him enough conceptual footing to be an equal participant in understanding.

He also embeds figures he can manipulate: dragging rocks around the garden and watching their coordinates move is how he comes to understand the isometric perspective. He notes that the document uses a feature Notion had just shipped, the ability to embed interactive HTML inside a page.

Only after that preparation does the explainer reach the code. A conventional diff presents edits as a collection of changed files ordered by filename, without explaining the conceptual sequence that connects them. What he calls a literate diff is structured as prose instead: it walks the changes in a sensible order with surrounding explanation and embedded code snippets. His judgment is that this is faster to review than the raw diff.

The finished result is an explainer packet. He still reads the code diff, but he always reads the packet first. He also points out the irony in it himself.

“It's beautifully ironic: AI turns an interactive activity into a static paper report I can focus on deeply”

His own caveat is that this leaves one problem in place: reading is hard work. He cites Andy Matuschak's formulation that "books don't work": it is too easy to fool yourself into thinking you did the reading when you did not actually retain or understand it.

His response draws on work by Matuschak and Michael Nielsen that embeds spaced repetition quizzes inside essays. Spaced repetition is a memory technique in which someone quizzes themselves on the same material at widening intervals. That work is Quantum Country, an online text with quiz questions placed directly in the prose.

Litt adapted the idea for code explainers. At the bottom of each explainer there is now an interactive quiz containing five questions about the change, and he tries to answer them. He applies a specific rule to his own work.

“My rule: I won't send code to others until I can pass the quiz, and I do the same when reviewing others' code.”

“A quiz is a speed regulator.”

That completes the first technique. Litt published the skill and offered it through a QR code during the talk, in two variants: one outputs HTML, one outputs a Notion page. They diverge on how the quiz is implemented. The HTML variant requires interactive multiple-choice questions that tell the reader whether they were correct and give feedback on click. The Notion variant requires each option to carry an explanation of why it is right or wrong, held in toggle blocks, with ❌ and ✅ marking those explanations.

When two people hold the same mental model, he argues, they can communicate efficiently because shared vocabulary evokes the same underlying images and structures. That common ground, in his words, is what lets people "jam and riff and have creative conversations". Without it, the same exchanges become much harder. He says he is interested in shared environments where teams build this understanding together, connecting the idea to the work of Notion, where he has already disclosed that he is employed.

As one concrete example, he describes features Notion has been shipping for humans and agents to work in the same environment. Claude and Cursor agents can now run in Notion, which he says is how he does much of his coding. When an agent produces a technical plan there, the plan appears by default in a collaborative page, so he can comment on it and discuss it with his team immediately rather than leaving each person to work with an agent in a separate silo.

“Thinking together, not alone!”