Julie Zhuo spent a weekend building an app most people would just buy. Frustrated with juggling eight terminal windows to run Claude and Codex agents across her projects, the designer and writer behind the newsletter The Looking Glass wrote a working version of a mobile agent-switcher in roughly 30 minutes, then spent several more days adding features nobody else would want: a sparkle count for active subagents, custom status verbs, a project she calls Mog Squad. Her argument, laid out in a long post on X, is that AI has made this kind of one-person software cheap enough to matter.
Zhuo’s case rests on a shift she says has already happened underneath most products without anyone renaming the category. Settings menus, she writes, were never real personalization: designer Jared Spool once collected Microsoft Word configuration files from several hundred users and found fewer than 5 percent had changed even one option. That is not evidence people dislike customizing software. It is evidence that toggles decided in advance by a product team rarely match what any individual actually wants.
What changes now, in her telling, is who gets to build the alternative. She sketches three tiers: fixed use-case apps, platform builders such as Shopify or Lovable that let users assemble something within guardrails, and what she calls the “anything builder,” a coding model with enough tool access to produce whatever a person imagines. Her own agent-manager and a second project, a Duolingo-style math and reading app she wrote for her kids using stories about their actual friends and hobbies, sit at that third tier. Both took her days rather than months precisely because a language model wrote most of the code.
The stronger piece of her argument is behavioral rather than technical. She points to a 2013 study of 145 ninth graders working through algebra problems. Half the group got the standard wording. The other half got the same problems recast around whatever each teenager happened to care about. Speed and accuracy both rose for the recast group, the students who had been struggling hardest gained the most, and the effect survived once the personalized wrapper was taken away again. That result predates any AI tool by more than a decade. It suggests the constraint on software customization was never demand. It was the cost of producing enough variants to satisfy it, which is exactly the cost curve coding assistants are now flattening.
Zhuo’s evidence is a sample of one. Every example in the essay is software she built for herself or her own children, and she is explicit that she is not claiming her apps beat the polished versions Anthropic and OpenAI ship, only that hers fit her better. Whether that generalizes to teams building for thousands of users, who cannot personally test every configuration the way Zhuo tests her own habits, is the open question her essay does not address.
For product teams, the actionable read is narrower than “AI ends generic software.” It is that any workflow a user repeats dozens of times a day, the kind Zhuo says friction compounds fastest in, is now a candidate for a coding agent to reshape without waiting on a vendor’s roadmap. Teams evaluating internal tools built on Claude Code or Codex should audit which daily workflows still route through rigid menus before a competitor’s engineers route around them first.
Julie Zhuo, writing on X on September 9, 2026, in an essay first published on her newsletter The Looking Glass.
