Can automate ≠ should automate

Recently I’ve been using ChatGPT’s Scheduled Tasks quite heavily, with 13 active tasks right now.

The feature itself is very simple:

Prompt + Frequency.

Tell ChatGPT what to do, then how often to do it.

It can be a one-off reminder, or repeat daily or weekly.

What really takes time to design isn’t the frequency, but:

which things are worth handing to AI on a schedule?

These days I roughly sort them into four kinds.

1. Repetitive searching, where AI can do a first pass of filtering for me

The most typical is job hunting.

I could of course reopen LinkedIn, company sites and job boards every week.

But what really takes time isn’t the search; it’s the chain after it:

finding new postings, deduplicating, checking location and visa requirements, judging whether AI is really a core responsibility, filtering out roles that obviously don’t fit, and then finding the few truly worth a look.

So now I have a Weekly Job Scan.

AI handles the search and the first round of filtering; I make the final call on whether to apply.

Weekend Discovery is similar.

Every week it finds events worth going to in Dubai on the weekend, checking times, weather, transport, prices, and whether it has recommended them before.

It isn’t that I can’t do these things; they’re just highly repetitive, and machines are well suited to doing them steadily.

2. Helping me come across things I wouldn’t have searched for myself

This kind is my favorite right now.

If I search myself, I’ll most likely keep searching AI, product, career and psychology.

But there’s a very real problem:

I can’t search for things I don’t even know exist.

So I made a set of tasks dedicated to “widening my input.”

• Essay Wildcard finds one article each week from a different field that’s genuinely worth reading.
• Culture Passport picks one way into culture each week: a film, music, architecture, an artist or something else.
• Open Door Scout finds talks, workshops, reading groups or communities I wouldn’t normally come across.
• Daily Companion also regularly includes things I didn’t ask about but that are worth knowing.

What I want AI to do here isn’t “understand me better and better,” but occasionally take me outside my own range of interests.

3. Looking back over a period regularly, instead of only answering the present

A lot of my thinking about work, learning, even life, already happens in conversations with ChatGPT.

Those conversations themselves have built up quite a lot of material.

So I started having it do regular organizing across time.

For example, Long View looks back over the past week each week:

• which things were real changes in fact;
• which judgments changed;
• what questions keep coming up;
• which things don’t actually have enough evidence yet to conclude.

There’s also a Writing Evidence Harvest.

It doesn’t conjure up three “topics” for me each week out of thin air; it looks back at what really happened this week, then judges whether there’s anything in it worth writing about further.

This kind of work suits LLMs very well.

It doesn’t need to form conclusions for me, but it can first put information scattered across different times back together.

4. Regularly checking things that are easy to overlook

This is a kind I only started using later.

For example, I’m developing Roleward now, and there’s a Weekly Audit.

Each week it mainly checks the changes from the past seven days:

• whether new inconsistencies have appeared between requirements, UX, data structures, code and docs;
• whether temporary patches are piling up;
• whether there are new reliability, cost or maintainability risks.
• I don’t need to re-review the whole repo every week.

What’s more valuable is checking whether recent changes left problems behind.

This kind of thing is a lot like regular housekeeping.

Important, but easily skipped again and again because it’s “not urgent.”

A scheduled prompt isn’t quite like an ordinary prompt

My earliest Daily Briefing was roughly:

“Every day, tell me the news on AI, product, career and health that I care about.”

Sounds fine.

But after running it for a while, it got more and more boring.

Because the system knew what I cared about, so it kept giving me more of the same.

Later I turned the prompt around:

don’t only start searching from my interest keywords;

actively cover different fields;

downweight topics that have been covered a lot recently;

don’t repeat the same topic several times in a row;

at least some of the content should come from directions I haven’t discussed at all recently.

Same with Weekend Discovery.

If it only says “recommend places that suit me,” it easily keeps recommending a few areas and shops it already knows I like.

So later I added:

check the recommendations from the past few weeks;

don’t repeat without a new reason;

keep at least one option slightly outside my usual range.

That’s also what I now think most needs designing in a scheduled prompt:

you can’t only write what you want it to do this time; you also have to think about what problems will appear after the task has run ten times in a row.

An ordinary prompt often only aims at whether this round’s answer is good.

A scheduled prompt also has to consider repetition, drift and long-term quality.

When I write a scheduled task now, I roughly consider this information

Not just:

recommend me some articles every week.

I usually also add:

Goal: what problem is this task ultimately solving for me?

Frequency: why does it need to run daily / weekly / monthly?

Input: what should each run draw on? The public web, recent conversations, previous results, or outside tools?

Filtering criteria: what’s worth keeping, and what should be filtered out?

History: does it need to refer to previous runs? What shouldn’t be repeated?

Exceptions: if there’s nothing really worth reporting this time, can it report nothing at all?

Output: what do I want to see in the end, rather than the model writing whatever comes to mind?

Boundaries: which things can only be analyzed and suggested, and must not be done for me?

Frequency is usually the last thing to be decided.

It isn’t limited to this round’s prompt either

If ChatGPT’s Memory / Chat History is turned on, a scheduled task can use the relevant historical context.

So some of my tasks say directly:

“Refer to relevant conversations from the past seven days.”

Or:

“Check the results of the previous few runs and avoid repeating them.”

But I don’t take this to mean “AI can browse all my history as it pleases.”

What context can be used still depends on account settings, the task environment and the data it can actually access.

Likewise, if ChatGPT is connected to outside tools like Gmail, tasks can use those sources where permissions allow.

The results themselves stay in ChatGPT, and you can also get task notifications through push / email and so on.

So scheduled tasks are more like:

having a prompt rerun at a set time, with whatever context it can access at that moment.

If you want to design your own scheduled tasks

These days I first ask four questions:

• Does this thing recur?
• Would AI doing it continuously be more consistent or more thorough than me starting fresh each time?
• Does it mainly need searching, organizing, comparing and checking, rather than making high-risk judgments for me?
• If I forget to do it myself, is it likely to keep getting put off?

Only if most answers are yes do I consider making it a scheduled task.

Then you can use this template directly:

I want to set up a scheduled task.

Goal: [what problem it solves for me over time]

Frequency: [daily / weekly / monthly / a specific time]

What to draw on each time: [the public web / recent conversations / previous runs / outside tools]

Filtering criteria: [what’s worth keeping, what needs filtering out]

Avoid repeating: [what not to repeat, and for how long]

Only tell me when: [the conditions under which output is really needed]

Output format: [what I want to see in the end]

Boundaries: [read only / don’t execute / which decisions stay with me]

If you don’t even know yet “which scheduled tasks should I set up,” you can first list the things that came up repeatedly over the last week or two, and have ChatGPT help you filter them, instead of asking it to generate a pile of tasks directly:

Below are things I’ve done repeatedly over the past week or two, often forget to do, or need to recheck regularly:

[fill in]

Don’t turn all of them into scheduled tasks directly.

First judge which really suit Scheduled Tasks. Prioritize things that are:

highly repetitive; where AI has a clear advantage in searching, organizing, comparing or checking; more valuable done steadily over time than whenever I happen to remember; where the cost of mistakes is manageable and the final judgment can still stay with me.

For each suitable candidate, give: why it fits, a suggested frequency, what rules the prompt should include, and what problems are most likely to appear over a long run.

Also say clearly which things aren’t worth making into scheduled tasks, and why.

My principle for myself now is simple:

AI handles ongoing discovery, organizing and checking. Judgment and decisions stay with me.