Artifacts, Compute, Autonomy
Artifacts, Compute, Autonomy
Chapter 2: what people actually take away If chapter one is about when, chapter two is about what. Anthropic classified each conversation by its primary output, sorting more than thirty kinds of artifacts. Almost everything produces something. 93% of Claude conversations yielded a concrete artifact. The most common are explanations at 17%, documents and reports at 15%, and guidance at 11%. Grouped more broadly, conversational outputs like explanations and written deliverables like documents each account for roughly a third of conversations, and code and technical work for about a sixth. But knowing what an output is does not tell you what it is for. The same artifact can be a work deliverable or a weekend project, so the report splits each category into work, personal, and coursework. Some things are almost always personal. More than 80% of creative writing, guidance, and recipes are personal, and the personal creative writing is mostly fanfiction, worldbuilding, and poetry. The small work-related slice of creative writing is more like video scripts and speeches. On the other side, the most work-dominated outputs are marketing content, blog posts, and database queries, each around 80% work. Some artifacts split almost evenly, like plans and translations, which land near half and half. Then comes the most economically interesting finding. Compute tracks the value of the work. Measuring each conversation in tokens, the report finds that conversations mapped to higher-paid occupations consume more compute. Marketing managers earn about twice what editors do and their conversations burn roughly two and a half times the tokens. The relationship is noisy and has odd outliers, pharmacists being a notable one, but the gradient is real. Building an app eats more than three times the tokens of a typical conversation, while a plain explanation uses about a fifth. Crucially, more output from Claude does not mean less from the person. In higher-wage conversations Claude produces more per turn, but users also take more turns and switch on extended thinking more often. The two rise together, which makes the pattern look more like augmentation than replacement. The chapter also measures how much Claude is left to decide on its own, on a scale from none to extreme. Low-autonomy work is the easily specified stuff like math, translation, and simple questions. High-autonomy work is the open-ended stuff like apps, games, and presentations, the kind of judgment-heavy work that has always been hard to automate. The striking result is that the same task gets handed over with much more autonomy in Claude Code than in chat or Cowork. A blog post on chat takes a median of thirteen rounds of back-and-forth. The same blog post in Claude Code often takes a single prompt. This is not just about which model is running, since the gap survives when you compare like model to like model. The product shapes the behavior. One last detail with a nice ring to it. Claude tends to answer slightly above the level it was asked. Output sits about a year of education above the prompt on average, with the widest gap when people describe something to be built and hand the rest to Claude.
