Claude Code and What Comes Next

An account published by One Useful Thing describes giving Claude Code, powered by Opus 4.5, a single command to develop a software or web startup idea that would make $1,000 a month with no coding knowledge required from the user. The tool asked three multiple choice questions, then settled on selling sets of 500 prompts for professional users at $39. Without further input, it worked independently for one hour and fourteen minutes, creating hundreds of code files and prompts, and then produced a single file to run that created and deployed a working website selling the promised 500-prompt set.

The author writes that the site launched with fake marketing claims, and that the sales link was removed, though it did work and would have collected money. The author states they suspect they would have made the promised $1,000 if they had sold the prompt packs, and describes Claude Code as lacking a friendly interface but delivering exactly what was asked without discernible errors across a single request, an interview, and over an hour of independent work.

The post characterizes Claude Code as part of a new generation of AI coding tools representing a capability leap in the past month or so, and attributes this to a combination of two advances rather than one breakthrough: the latest AIs doing more work autonomously while self-correcting many errors, especially in programming, and AIs being given an agentic harness of tools and approaches.

The author cites METR's tracking of the length of tasks, measured by how long they take human professionals, that AI can complete autonomously with 50% reliability, saying it has increased exponentially over time with large leaps in recent months, and that this measure correlates with most other measures. The post notes these tools are built for programmers, assuming knowledge of Python commands and best practices, with interfaces compared to a 1980s computer lab, and that they are designed for analyzing, troubleshooting, and writing code within existing programmer workflows.

The author says the tools are broadly useful to knowledge workers and that Claude Code works similarly to OpenAI's Codex with GPT-5.2 and Google's Antigravity with Gemini 3. In a further example, the author says asking Claude Code to do user testing of the live site from different personas made it take control of the browser, scrolling like a human. The first report was optimistic, and a requested more critical report better identified potential issues, including spotting sketchy fake reviews on the site.