David Wilson catalogued 16+ projects he spun up with AI tooling and arrived at a blunt conclusion: cancel the subscription.
The pattern he describes is specific and familiar:
- Start small — a Claude session begins with “write a quick script for X”
- Scope explodes — one hour later, the result is neither a quick script nor a solution to the original problem
- Repeat across projects — 16+ half-built things, none maintained, most abandoned
Wilson’s sharpest line: the technology is “a thermonuclear ADHD amplifier.” He reports seeing the same behavior in every adult friend who uses coding agents — people running three screens simultaneously on unrelated projects they have no commitment to finishing.
Simon Willison, who flagged the post, adds his own recognition of the problem. Coding agents can take a vague idea to a working solution complete with tests and documentation — something that looks like a carefully considered project. That’s the trap. The friction that used to kill bad ideas before they consumed hours is gone. Every impulse gets a prototype.
This isn’t a tooling critique. It’s a workflow critique. The agent did exactly what was asked. The problem is that “what was asked” shifted twelve times in an hour because the cost of asking dropped to zero.
NOTE
Wilson’s proposed fix isn’t better prompting or stricter project management. It’s removing access entirely — cancelling the subscription to reintroduce the friction that filtered out low-value work.
The post resonates because it names something most heavy agent users have felt but not articulated: the gap between output volume and problems actually solved is widening. More code gets generated. Fewer projects ship.
Wilson’s 16 unfinished projects are the metric that matters here — not tokens consumed, not lines generated. If your agent-assisted project list is growing faster than your shipped-project list, the subscription might be the problem, not the solution.