Keep AI Side Projects Weekend-Sized
Audit coding-agent subscriptions with a three-question stack check, then cut the project back to something a person can use on a weekend.
Your coding-agent subscription is doing the work, but the weekend build is still a prompt file that describes the next prompt. Steve Yegge's Gas Town was meant to be reusable, but he ended it and conceded that, after paying substantial monthly fees for coding-agent subscriptions, the only thing he built with it was the system itself.
That is the failure mode: the agent gets busy, the stack grows, and the thing a person can use stays missing. The three-question check keeps the project weekend-sized.
Run the check before the next prompt
Keep the three questions visible, because the drift is gradual: one extra prompt, one extra orchestrator, one extra subscription tier. Each item should take a minute, and each answer should change what you do next. Use it when the project feels busy but not closer to done.
- Is the agent output the product, or is it scaffolding for the next agent run?
- What did you ship that is not agent configuration?
- Is subscription spend outgrowing the project?
If the answers point to more tooling, stop adding features to the tooling. The project is still yours, but the scope has drifted past a weekend. If you cannot answer in a minute, the answer is probably no.
Scaffolding is the first failure
The Opus 4.7 coding model introduced a recurring just-two-more-things behavior that stopped convergence, so the work kept reopening instead of finishing. That pattern is easy to miss when the agent is busy. It writes plans, edits files, and suggests the next refactor. The output looks like progress, but if the agent is mainly improving the agent, the project has stopped producing value a person can use. Cut the stack until the next run produces something outside the stack.
The agent can be productive and still be building the wrong thing. It may generate clean code, useful notes, and a tidy architecture. None of that matters if the next feature a person can use is still missing.
Configuration is not a shipped result
The second question asks for a result, not just configuration. Prompt files, orchestrators, and cleaner sets of instructions are all useful, but none of them is the product. Dan Luu found ultra-vibed orchestrators unreliable because they could not reliably finish tasks. Reliability matters more than cleverness when you have a few hours a week.
If the only result is a better prompt, ship a smaller result someone can use. A weekend-sized project should leave behind something usable, even if it is rough. The agent can help, but the shipped result should not be a more elaborate way to ask the agent for help.
The shipped result can be small: a page that answers a question, a script that saves time, a report a person can read. It does not need to be polished. It needs to be outside the agent's configuration.
Spend sets the scope
The third question turns spend into scope. If your agent bill is growing faster than the value someone can use, the project has outgrown a weekend. After Claude Fable 5 became available, Steve Yegge resumed full-time work on Wyvern, a video game he started in 1996. Wyvern development consumed API tokens equivalent to $87,000 per month, roughly 69 billion tokens in July. Yegge offset that token burn with $200 Max accounts, leaving an out-of-pocket cost of about $2,800 per month. A side project should not need a special account stack to stay alive.
Set a hard limit before the next weekend. When the bill approaches it, remove one agent, one subscription, or one extra layer of coordination. Cost should match the size of the thing you are trying to ship. The limit should be boring. Pick a number you can defend, then compare it to the value you can name. If the value is smaller than the bill, cut the stack before the next weekend. If the bill stays far below the limit, you may keep the stack, but only while the project is producing something a person can use.