I build an app that uses AI to help schedule your day. You’d think I’d have the AI planning workflow figured out. I don’t, entirely. Here’s an honest account of what I actually do, what works, and where the wheels fall off.

My actual setup

Every morning I open a scratch doc and paste in three things: a rough brain dump of what’s on my plate, any hard deadlines or meetings for the day, and a short note about how I feel (slept poorly, post-meeting fatigue, sharp and ready, something like that).

Then I paste that into a prompt that asks Claude to draft a prioritized plan for the day. Not a fixed schedule with time blocks, just an ordered list with a short reason for each item.

That’s it. No elaborate system prompt, no fancy chaining. A dump in, a draft out.

I’ve iterated on the prompt a few times, but the format has stayed basically the same for months. The real work isn’t the prompt. It’s what I do after I read the output.

Where it genuinely helps

Triage when my head is full. When I’m carrying ten things and can’t figure out what to touch first, externalizing the list and having something sort it is useful. Not because the AI is smarter than me about my priorities, but because seeing a proposed order breaks the paralysis. I read it, disagree with two items, agree with the rest, and move on. That reaction is information.

First-draft plans for vague tasks. If I have something on the list like “figure out onboarding flow,” the AI will ask clarifying sub-questions or just generate a set of concrete sub-steps. Half of them are obvious, a few are wrong for my context, but one or two are angles I hadn’t thought of. That’s a reasonable return for thirty seconds of work.

Breaking tasks apart. I’m prone to listing things like “fix authentication,” which is actually four separate things. The AI is good at noticing that and splitting it out. It does this reliably even when I can’t be bothered to do it myself.

Where it fails

None of these show up in demos.

It doesn’t know what it doesn’t know about my context. I might list “write landing page copy” as a task, and the AI will schedule it as a medium-effort item. But if I’m waiting on a design review that blocks it, or if I had an argument with a collaborator yesterday that makes me not want to touch it today, the AI has no idea. It reasons from what I give it, and I often give it an incomplete picture.

Confident wrong suggestions. This is the one I’ve learned to watch for. The AI will sometimes suggest I tackle a complex technical task “while I have the most energy in the morning,” but that assumes a morning-peak pattern that doesn’t apply to me. Or it’ll tell me to “batch the shallow tasks in the afternoon,” which sounds smart but ignores that my late afternoons are often broken up by async messages. The suggestions sound structured and authoritative. They’re not always wrong, but they’re never calibrated to me specifically.

No visibility into energy or mood. I include a mood note in my prompt, but the AI treats it as one data point among many. It doesn’t have a model of how a bad night’s sleep actually degrades my ability to write clearly versus debug code. It can’t tell the difference between “tired but functional” and “tired and will make errors.” I can tell that difference. The AI is guessing.

The plan is fresh the moment I generate it and stale an hour later. Something urgent comes in, a meeting runs long, I go down a rabbit hole. The AI’s tidy plan is now a fiction. It has no ability to adapt in real time. I’m back to manual triage, which is fine, but it means the planning step has a short shelf life.

The guardrails I use

After enough friction from taking AI plans too literally, I’ve settled on a few habits.

I treat every output as a first draft that needs a human edit. I read it, mark up the two or three things that are clearly wrong for today, and reorder the rest. This takes maybe two minutes. It’s fast precisely because the AI gave me something to react to rather than starting from a blank list.

I keep a hard rule: nothing goes on my actual to-do list until I’ve decided it belongs there. The AI suggests; I decide. This sounds obvious but I’ve broken this rule when I was tired and just accepted the output wholesale. It always went badly.

I also don’t use AI for real-time re-planning. If the morning plan falls apart, I do that myself. Asking the AI to re-plan mid-day with all the accumulated context of what actually happened is more friction than it’s worth.

What Flow does with this

Building Flow, I’ve had to think carefully about where AI assistance is worth the complexity and where it’s just theater.

The app uses AI in one specific place: helping you draft a plan for the day based on your task list and how you’ve rated your energy that morning. The output is a suggested schedule, not a locked one. You can drag things around, swap blocks, cut items entirely. The AI draft is a starting point, not a commitment.

What Flow doesn’t do is let the AI decide anything for you. It won’t auto-schedule and lock your calendar. It won’t silently reshuffle things when a block runs over. Those decisions stay with you, because you’re the only one with the full picture.

The reasoning behind this is explained in more depth in the post on energy-aware scheduling and the idea behind Flow, but the short version is: an AI that acts as a confident planner without knowing your energy state is solving the wrong problem.

The AI planning step in my own morning routine works because I stay in the loop. I use it to save five minutes of mental effort and to catch the task I was about to forget. I don’t use it to tell me how my day should feel or what I’m capable of right now. That’s mine to know.

If you’re building something similar for yourself, the useful test is: are you making a decision based on the AI output, or are you just handing the decision to the AI? The first is a tool. The second is abdication with extra steps.

If any of this resonates and you want to try the approach with Flow, you can sign up for early access here. I’m building it in public, so feedback from early users shapes what gets built next.