A thirty-page site with a CMS, structured data, a full content layer and a case-study system used to be a month of work. I built one in days. The claim sounds inflated until you look at where project time actually went, which was never typing.
Three things consumed it: starting from nothing, switching between unrelated kinds of work, and waiting for someone else to review or revise. AI tools removed most of all three. Here is the actual workflow.
The blank page was always the expensive part
Starting a new component, a new page type, a new content model - the cost is not the typing, it is the deciding. What shape should this take, what does it need to handle, what is the structure.
A model produces a complete, plausible first version in seconds. It is frequently not the version I ship. But editing something concrete is dramatically faster than deciding in the abstract, and the first draft surfaces the questions I had not thought to ask.
This alone is a large fraction of the speedup, and it is the part people underestimate because it does not look like work being done.
Each tool for what it is actually good at
Claude for reasoning through structure and reviewing whole files. When I need to decide how a content model should be shaped, or check whether a file has drifted into inconsistency across several edits, this is where that happens.
Cursor for changes inside the editor with project context already loaded. Small, surgical edits across files where the surrounding code matters more than the explanation.
Codex for well-specified isolated tasks - a script, a transform, a utility with clear inputs and outputs and no dependency on the rest of the codebase.
Using the wrong one is the most common source of frustration. Asking for an architecture decision in a tool optimised for inline edits produces a confident answer with no context behind it.
Context switching stopped being expensive
The old rhythm was batching: design for a day, build for a day, write copy for a day, because switching cost twenty minutes of reorientation each time.
That cost collapsed. I can design a section, build it, write its copy and move on, because picking up each thread no longer means rebuilding the whole mental model from scratch - the tools hold the context.
The practical effect is that work stays coherent. Copy written while the layout is in front of you fits the layout. Copy written three days later on a different day fits whatever you remember of it.
Where I spend the saved time
This is the part that determines whether the speedup produces good work or merely fast work.
I spend it on judgement: does this structure actually serve the user, is this copy honest, does this section earn its place. And on the invisible layer - what happens when the CMS is unreachable, when an image is missing, when someone submits the form twice.
I found a real example of the second while building my own site. A build that could not reach the CMS silently produced a complete, valid, empty website. Every page built, nothing errored. Deployed, it would have replaced a working site with nothing. Now it throws.
That is the trade. Generation handles the implementation. The saved hours go into the two things it is worst at: whether the thing is right, and what happens when something breaks.
What 10x does and does not mean
It does not mean every task is ten times faster. Writing a tricky algorithm is maybe twice as fast. Debugging something subtle is sometimes slower, because a confident wrong explanation costs time to disprove.
What is ten times faster is the bulk of real project work: scaffolding, repetitive structure, content population, boilerplate, the second through twentieth variation of something you already decided. That is most of a project by hour count, even though it is the least interesting part.
So the honest version of the claim is: the boring eighty percent collapsed, and the interesting twenty percent got more attention than it used to. That is a better outcome than the headline number suggests.
Questions
Is 10x faster with AI a realistic claim?
For the repetitive majority of project work - scaffolding, boilerplate, repeated structure, content population - yes. For genuinely hard problems it is closer to twice as fast, and subtle debugging can be slower. Since the repetitive part is most of a project by hours, the overall effect is large.
Which AI coding tool should I use?
Use several for what each does well: a reasoning model for structural decisions and whole-file review, an in-editor tool for surgical changes with project context, and a task-focused tool for isolated, well-specified jobs. Most frustration comes from using one tool for all three.
Does working faster with AI mean lower quality?
Only if the saved time goes nowhere. The speedup is in implementation, which is not where quality is decided. Redirect the saved hours into judgement and failure-case testing and quality goes up, because those are exactly the areas generation is worst at.