My old job feels very distant now
Short summary
A developer walks through how their daily coding work changed across ChatGPT, Cursor tab completion, Claude agents, and Opus 4.5 through 5.5, ending with a model building a working notes app for iPhone from a single prompt. It reads as a personal timeline of AI tools moving from a better search box to something that handles most of the implementation.
- Productivity rose from a modest gain to roughly 3x (measured over a few months of merged PRs) once the Opus 4.5/6 models needed far less step-by-step prompting.
- The developer's role shifted from writing and reviewing every line to focusing on architecture, business logic, customer data, and irreversible changes, scanning the rest.
- A single prompt produced a basic working iPhone app that imported all their Electron notes, using about 40% of a $20 subscription's 5-hour limit.
Full summary
The author traces how their coding workflow shifted from ChatGPT as a better Google, to copy-pasting Ruby specs they hated writing, to tab completion inside Cursor (a VS Code fork), and finally to agentic prompting on Claude. The early ChatGPT phase mattered most for a computer vision side project in Python, where they knew nothing about the domain going in. They ran code in Jupyter, tinkered, and learned the APIs as they built. Hallucinations were a useful signal: they clustered at the edge of what the model knew, which told the author where to dig into poorly documented or project-specific behavior. Their stance is that this was about speed and less SEO spam and bad docs, not giving up control.
The agentic stage is where the article gets concrete. The author describes writing long, precise prompts that name specific files and existing patterns to mimic, like telling the model to copy an existing component and follow community idioms with stubs instead of mocks. Prompting went from a trick to the main way they worked. They kept the codebase fully in their head and reviewed every line, and they say productivity rose noticeably because they no longer had to write code they disliked. The real bottleneck became PR review. Their answer was to lean on standard libraries and established patterns so that most of a diff could be scanned rather than read closely, and to focus human attention on business logic, customer data handling, and irreversible changes like database migrations and encryption.
The jump with Opus 4.5/6 cut how much hand-holding the author needed and, by their own measurement over a few months of merged PRs, roughly tripled output, with better feature polish. Later models improved code quality and reduced corrections, but verbosity and cost became the irritation. The author briefly switched to ChatGPT, which they found nearly as good for their use and cheaper and less wordy, though they distrust OpenAI and went back once Anthropic's model caught up. With Opus 5.5 the work changed again: they now aim more ambitious side projects, produce plenty of throwaway output, and say they've begun planning around QA rather than code review, which is the shift they're still working out.