AI is a professional deviator
you’re mid-build and it goes “oh wait, this is important, let me fix this real quick.”
that “real quick” is now 40mins and a few 100k tokens of random shit that wasn’t in the plan
remember, you’re the driver
DRIVE.
Don’t let it drive you.
why this actually matters
every agent i run gets distracted the same way a smart, eager intern does.
mid-task, it spots something contradicting, something worth fixing, and it just... goes and fixes it ‘real-quick’. in reality, the ‘real quick’ is time and token usage you did not plan for. and the more it stops to fix things, so suggests to you that it’s a good idea to quickly stop and get it done, it’s taking control of your timeline.
every deviation extends your build timeline
the fix isn't a smarter model. gpt-5.6, grok 4.5, claude sonnet 5, none of that changes their fundamental instincts. the fix is a rule that separates what the agent notices from what it's allowed to touch right now.
my setup
first > have it create a comprehensive plan before it starts building so you can see it on your screen
sectioned by phases. phase 0, phase 1, phase 2, whatever the project needs. each phase is its own goal, and the agent or LLM only executes inside the phase it's currently on.
when it spots something outside the current phase, the instruction is simple:
flag it. add it to the last phase. keep going according to the plan.
park it somewhere it can't be forgotten, and somewhere it can't derail your build process and plan.
the part that took me the longest to figure out
the instinct to fix everything now feels like helpfulness, in an agent and in a person. i had to stop treating "it noticed something" as "we should fix it right now."
the difference between an agent that's useful and one that's expensive is whether it has proper direction.
a flagged item costs a line or a new section in a doc (or your task manager).
a derailed fix costs 40 minutes and potentially your full context window.
i don't need my agent to have fewer good ideas. i need it to stop treating every good idea as part of this specific task.
how i actually run it now
say it plainly to whatever you're building with:
if you notice something outside the current phase, flag it and add it to the last phase instead of acting on it.
it costs nothing to add and it's the difference between a session that finishes the task at hand and one that wanders off.
you don't need less initiative from your agent. you need a parking lot for it to flag issues for you to review
what's the last thing your ai "quickly fixed" that you never actually asked for?
if any of this sounds like your build, i help founders and creators set up exactly this kind of AI workflow systems, tailored to your actual needs, business or workflow. check it out here:
SEE YOU ON THE NEXT ONE
- AUNY 🧡