They can save a great deal of time, but only if used with a tight leash. You have to know what you are doing in order to give it the instructions and not let it either make guesses or get out of control. For the first part of that, I wrote about it here. In short, flip the script and have it let you know about any ambiguities. As for the control aspect, it can easily lose track of context and revert to default behavior, which is often not what you want. So, you have to constrain it. For example, you may want to add this to a prompt: “Perform analysis only. Do not make any changes yet.” Hopefully, this can let you review and correct what it may be planning to do before it does it. Once you’re ready, add a bit more to the prompt: “Proceed with just this set of changes. Do not add, change, or delete any other code.” Further, as the set of changes grows in complexity, break it down and have it work on and track each set of changes as it goes.
Another critical aspect, keep your repo commits up to date. Update the repo at each small point of stability. Then, when the agent does lose control, it is easier to revert and start that set of changes over again. Some may recall the Dilbert cartoon about ‘Write a minivan’. Sometimes, the agent may want to do this also. You may pay it to write something and, after ignoring it for a while, pay it again to fix what it just did. So, when you run agent prompt instructions, keep and eye on it and abort it if it gets off track.
I wrote about using AI/agents for greenfield work here. In this initial Elixir/Phoenix/LiveView project, more than 90% of the codebase was generated. It has since been through two AI/agent assisted refactors, one for adding full Stripe product catalog synch and another for adding user usage metrics and cost tracking. The current state of the project is a mostly working app. But, as a solo-preneur, generating code is the easy part. Not all tests can be automated. You still have have to manually test functionality to ensure it is working correctly and not just passing tests, which may or may not be correct. That human time is still one of the slow parts of any sufficiently complex project.
To the flow in the post, I have since added use of Cursor for smaller sets of refinements. Those 500 premium requests per month helps to reduce the Claude Code costs.
But it really gives me a visceral reaction every time I hear it and it doesn’t seem to be getting any better.
