“Daily” was doing some ambitious work in that sentence.

I’m a writer, publisher and entrepreneur. I expect a certain amount of chaos. There are manuscripts, emails, projects, follow-ups and the occasional administrative mystery that arrives disguised as a perfectly ordinary Tuesday.

Still, 156 items suggested that my productivity system had developed career ambitions of its own. It was building a department. Soon it would need an org chart.

Some of the work was already underway. Some items overlapped. Some were reminders about other reminders. Tasks that belonged together had become separate obligations, each carrying the emotional weight of an unfinished thing.

I had asked for help getting work done. I was getting very good at maintaining evidence that work existed.

I use AI, and I remain a big believer in what it can do. It helps organize ideas and get thoughts onto a page. Technology that makes expression and work more accessible matters to me.

That is why I want to be honest about the less glamorous part: an assistant can generate work faster than it can finish it.

Ask it to help with a project, and it may produce a plan. Ask it to improve the plan, and it may produce a checklist. Ask it to make sure nothing gets missed, and the checklist may develop subcommittees.

Every item can sound reasonable on its own. Collectively, they can turn help into a second job.

The problem starts with an easy confusion: identifying the next step feels like progress.

Sometimes it is. A clear plan can get a stuck project moving. But a plan also has a comforting property: it can be created without confronting the hardest part of the work.

A draft can sit there looking polished. A calendar can fill up with preparation. A task can be checked off because someone wrote instructions for completing it.

Meanwhile, the email still needs sending. The application still needs submitting. The client still needs the finished file.

A productivity system earns its keep when those things happen.

After the 156-item surprise, I started insisting on a more useful question: What would “done” look like to the person who asked for this?

For an article, that could mean a finished draft, if a draft was requested. If the request was to pitch it, “done” includes an actual pitch. If the request was to publish it, a document called FINAL_FINAL_REALLY_FINAL does not quite complete the journey.

The distinction sounds obvious. It gets surprisingly slippery once work passes through several tools, checklists and assistants.

Here are the changes I’m making.

First, I want a daily list that reflects available capacity.

For assistant execution, I’m aiming for roughly 25 meaningful entries on a weekday, while preserving real deadlines and necessary daily operations. The number is a working limit for my system, rather than a productivity prescription for everyone.

Its purpose is to force a choice. If everything is due today, the list has stopped helping me decide what to do today.

Flexible work can live in a backlog with a reasonable checkpoint. Waiting for someone’s response deserves a waiting status. A genuine deadline deserves space that has not already been occupied by 40 vaguely urgent reminders.

Second, related work needs a home.

A campaign may involve research, outreach, tracking and follow-ups. Those requirements still matter. Putting them under one clear execution item can preserve the work while reducing the number of separate things shouting at me.

The details should remain available when they are useful. I don’t need every detail to greet me at the top of the stairs.

Third, I want proof attached to completion.

If an assistant says an email was sent, I should be able to see it. If it says something was published, I need the live link. If it says a file was delivered, I need the usable file and the delivery record.

That proof also helps prevent duplicate work. A system that remembers what actually happened can stop proposing the same action with slightly different punctuation.

Fourth, a blocker needs to be specific.

“Unable to complete” gives me another mystery to manage. “The form requires an answer only you can provide” gives me a decision. “This browser session cannot access the account” identifies an access problem.

The next question is whether anything else can move forward. One missing answer should not freeze every unrelated part of a project.

Finally, I want the system to count outcomes.

How many pitches actually went out? Which replies need an answer? What has been accepted? Which deliverables reached the client? What remains unfinished?

I care about those numbers because they describe the work I was trying to accomplish. A magnificent list of possible actions describes the system’s imagination.

These changes require judgment. Some days have more real deadlines. Some tasks take minutes; others need uninterrupted hours. A short list can still be unrealistic, and a long checklist can be useful inside a complicated project.

The goal is a system that helps me see the difference.

AI can be extraordinarily helpful here. It can reconcile records, find overlaps, preserve context and handle routine steps. But it needs instructions that reward finishing, checking and communicating honestly about what remains.

It also needs permission to stop manufacturing another plan when the next useful action is already clear.

I still want an assistant. I still want technology to make work easier and possibilities bigger.

I would just like my to-do list to stop recruiting.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.