Cancelling Your Claude Subscription as a Productivity Solution

David Wilson published a post this week that is already circulating widely: he lists over 16 projects he started with AI tools without planning them. The pattern he describes is recognisable: you open a Claude session to "write a quick script", and an hour later you have a half-finished project, the original problem unsolved, and three new tabs open.

Simon Willison picked up the thread on his blog on May 31st and added that he experiences it too: code agents are particularly skilled at taking you from a vague idea to a partial implementation without you consciously deciding to start a new project.

Wilson's diagnosis: "thermonuclear ADHD amplifier"

Wilson's phrase is harsh but precise. He describes watching friends work across three screens on unrelated projects, projects they will probably never maintain and have so little commitment to that the time invested simply evaporates. It is not a criticism of technology in the abstract; it is an observation about how the low cost of starting something with an agent removes the friction that once acted as a natural filter.

Before, setting up a project meant configuring your environment, writing boilerplate, making architectural decisions. That cost meant you only started what you truly wanted to finish. With Claude Code and a model like Sonnet 4.6 or Haiku 4.5 generating scaffolding in seconds, that filter disappears. The result is not greater productivity; it is more work surface without greater capacity to close it.

Why this matters beyond individual cases

This kind of reflection has practical relevance for teams that have adopted code agents in their workflow. If one person can accumulate 16 unfinished projects in a few months, a small team can generate a debt of context and maintenance that becomes difficult to manage.

The problem is not exclusively about personal attention. It is also structural: agents are very good at starting and notably worse at closing. They generate functional code for the happy path quickly, but debugging, documenting, integrating tests, and shipping something to production requires a kind of sustained attention that the "chat session" format does not favour.

Willison notes that the solution Wilson proposes, cancelling the subscription, is more a metaphor than literal advice. The real point is that using these models without prior discipline about what problem you are solving and when you consider it solved is a sure way to accumulate debt without assets.

What to do about it

Some practices that have emerged in the community to counteract the effect:

• Define your exit criteria before opening the session. What needs to happen for you to close the chat and mark the task complete?
• Separate exploration from execution. One session to understand the problem, another to implement it. Not mixing them reduces drift.
• Treat agent-generated projects as prototypes by default, not production code, until you deliberately decide otherwise.
• Review your project history monthly. If more than half are inactive, the problem is not the tool; it is your prior selection process.

None of this is new in project management. What is new is that agent speed makes the lack of discipline have visible consequences much sooner.

---

From our perspective at ElephantPink, the conversation Wilson and Willison open is more useful than most AI productivity tutorials: it focuses on the hidden cost of ease, which is exactly what few want to discuss when they are selling subscriptions.