Useful ChatGPT Prompts I Reuse Every Week

The results page for useful chatgpt prompts sorts by domain. Fifteen categories on one site, ten lists of ten on another, a hundred prompts here, two hundred tested there, and the top thread's favourite is "Write a gratitude list for 5 things you appreciate in your life." Useful, on that page, means covers a lot of situations.

I'd define it differently. A prompt is useful when you reuse it, and you reuse it when three things are true. The moment it serves comes round again, the input is different every time, and the output can be checked against something you already have. Seven prompts below pass that test across an ordinary week, one for each moment that repeats. The Monday plan, the long document, the reply, the decision, the error, the table of numbers, and the Friday review. They're shown in full with the structure explained, and what each one tends to produce is my judgment rather than a log.

Two of the seven can be put on a schedule. The other five can't, and the reason is the most useful thing on this page.

The Reuse Test

Three questions, asked of any prompt before it's saved. Does the moment recur? Is the input fresh each time, or is the model supplying it? Can the output be checked against something that isn't the model?

Here's the test run on prompts from the results page. The verdicts are mine.

The middle column does most of the work. A prompt with no fresh input is a prompt the model could have run without you, and what comes back will be the same for everyone who pastes it. Deep, crazy, life-changing, the adjectives in the related searches all describe prompts where the model supplies the content. The ones that change anything are the ones where you do.

The Seven

Each prompt is in a code block, ready to copy. Below each one is what the structure is doing and what the output tends to look like. Where I say "tends to", that's judgment from use, not a measurement.

Monday, The Plan From A Dump

Below is everything on my mind for this week, in no order. Do not add tasks. Sort what I wrote into three groups: things with a date attached, things blocked on someone else, and things blocked only on me. Inside each group, keep my wording. Then name the one item that, if it were done by Friday, would make the most other items easier or unnecessary, and say why in one sentence. Do not write a schedule. Do not motivate me.
Dump:
[PASTE]

No role, because the job is sorting and a role only adds vocabulary. The input is the dump, typed fast and badly. "Do not add tasks" is there because the current models fill gaps by default. OpenAI's guide for GPT-6 Astra describes a model that "tends toward detailed, formatted responses", and a plan is the format it most wants to write. "Keep my wording" is the check. If the output rewrites an item, you'll notice, and if you can't find your item in the output, something was dropped.

The third group tends to be the shortest and the least comfortable one to read.

The Long Document Into Asks

Read the document below. I need to know what it asks of me, not what it says. List every sentence that requires an action, a decision, or a reply from me, quoted exactly, with the deadline if one is stated and "no deadline stated" if not. Then list every number or date in the document that I should verify before acting on it. Do not summarise the document. If nothing in it asks anything of me, say so in one line.
Document:
[PASTE]

"Quoted exactly" is the whole design. Every line the model returns can be found in the document with a search, so checking the output takes seconds and a hallucinated ask fails the search. "No deadline stated" is the out. Without it, the model tends to infer a deadline from tone, and an inferred deadline looks the same as a real one. The summary is banned because a summary is what you'd get without a prompt at all, and you'd have to reread it to find the asks anyway.

The Reply From Three Points

Write a reply to the message below using only the points I give you. One point per paragraph, in the order I gave them, and no point I didn't give. No greeting line, no sign-off, no "thanks for reaching out", no exclamation marks. If one of my points doesn't answer something the message asked, add a single line after the reply, outside it, telling me what's still unanswered.
Message:
[PASTE]
My points:
1. [POINT]
2. [POINT]
3. [POINT]

The points are yours. The sentences are the model's. That division is what makes the reply checkable, since you can read it against three lines you wrote and see whether anything was added. The line outside the reply is the second check, and in my judgment it's the line that earns the prompt its place, because the thing the message asked that you forgot to answer is the thing you'd have been asked again on Thursday. If the reply also needs to sound like you, the voice spec from chatgpt prompts for writing goes at the top and this prompt sits under it unchanged.

The Decision, With What Would Have To Be True

I'm deciding between the options below. Do not recommend one. For each option, write the two or three things that would have to be true for it to be the right choice, as plain statements I could check. Then, for each statement, say whether anything in what I've told you shows it to be true, false, or unknown. Do not add options. Do not write pros and cons.
Decision:
[ONE SENTENCE]
Options:
[OPTION A]
[OPTION B]
What I know so far:
[PASTE]

"Do not recommend one" removes the least useful part of the default output, which is a confident pick based on whatever you happened to paste. What's left is a list of statements and, beside each, a verdict of true, false, or unknown. The unknowns are the output. They're the list of things to go and find out before deciding, and the first time you run this the column tends to be mostly unknown, which is a fair description of most decisions on a Tuesday.

The Error, Before The Fix

Below is an error and the code or command that produced it. Do not fix it yet. Tell me, in order: what the error message literally says, in one sentence; the three most likely causes, from most to least likely, each with the single thing I'd check to confirm or rule it out; and which of those checks is fastest. Stop there. I'll tell you which check came back positive.
Error:
[PASTE]
Code or command:
[PASTE]
What I've already tried:
[PASTE]

I keep a lot of projects in one repository, and the error I'm reading on Tuesday is often in a language I wasn't writing on Monday. The first sentence, what the message literally says, is the one I read. "Do not fix it yet" and "Stop there" are both in the prompt because one of them isn't enough. Left to itself the model tends to diagnose and then fix in the same breath, and the fix is usually for the most likely cause whether or not it's the actual one. The "already tried" slot is fresh input with a job. It stops the first suggestion being the thing you did an hour ago.

The Number Table Into Changes

Below are two exports of the same table from two dates. For each row that appears in both, report only the rows where the value changed by more than [THRESHOLD], with the old value, the new value, and the difference. Then list rows that appear in only one export. Do not interpret the changes or suggest causes. Do not round. If the columns don't match between the two exports, stop and tell me which columns differ.
Export from [DATE 1]:
[PASTE]
Export from [DATE 2]:
[PASTE]

The threshold is the slot that stops you reading noise, and it's the one you'll adjust most. "Do not round" is there because you'll check the arithmetic on any row you act on, and rounding hides the mistakes you're checking for. The column check at the end catches the most common failure, which is two exports that aren't the same table.

This is the prompt I'd point at a Search Console export for the site of mine that averages position 47, where the average hasn't moved in months and the rows underneath it have. Check the subtraction yourself on anything you'll act on. The model can miscount, and the prompt is shaped so that checking is cheap.

Friday, The Review Against Monday

Below is the plan from Monday and a dump of what actually happened this week. Do not grade me. For each item in the plan, mark it done, moved, or dropped, using only what the dump says, and quote the line in the dump that supports the mark. Items in the dump that weren't in the plan go in a separate list called "unplanned". Then answer one question in one sentence: was the item named on Monday as the one that mattered most done, and if not, what took its place. No advice.
Monday's plan:
[PASTE]
What happened:
[PASTE]

Monday's output is Friday's input, which is why the two are a pair. The plan is the thing the review is checked against, and the quoted line is how you check the review. "Do not grade me" and "No advice" do the same job as Monday's "do not motivate me". The model's default register for a weekly review is a coach, and a coach is the last thing a Friday needs.

The unplanned list tends to be longer than the plan. Read it as the record of where the week went, which is the one thing Monday's plan couldn't have told you.

The Reuse Ledger

One row per prompt, with the two columns the reuse test cares about and a fourth that the next section explains.

Which Ones Can Be Scheduled

ChatGPT can run a prompt on a schedule. The help center article on scheduled tasks, as archived in September 2026 since the live page refuses automated readers, says tasks can be one-time or recurring, that they're available to eligible Free, Go, Plus, Pro, Business, Enterprise, and Edu users, and that you create one by asking ChatGPT to do something and setting the schedule under Scheduled. Notifications go out by push or email from Settings, then Notifications.

The limits decide what's worth scheduling. Active tasks are capped at 3 on Free and Go, 5 on Plus, 10 on Business and Edu, and 15 on Pro and Enterprise. Free accounts get a recurring task at most once a day, delivered in a window such as morning or afternoon, and exact times need a paid plan. Tasks don't work with voice chats or GPTs. And the line that matters most for this article is short. A task created inside a project can't access the project's files.

So a scheduled task can carry the prompt. It can't carry the input. The Monday plan and the Friday review are the two I'd schedule, in the narrow sense the page supports. The task fires at the right time, in its own chat, with the prompt text and a notification, and the dump or the week gets pasted when it does. If you'd hoped to keep Monday's plan as a project file and have Friday's task read it, the page says no. Paste it.

The other five can't be scheduled at all, because their input arrives when it arrives. A document lands, or an error appears, and the prompt runs then. The one route on that page where input shows up on its own is event-triggered tasks, which respond to Gmail, Slack, or GitHub activity on Plus, Pro, Business, Enterprise, and Edu plans, and that's a different article.

What Memory Carries Between Runs

The tempting shortcut is to skip the paste and let memory carry it. The Memory FAQ, in its August 2026 snapshot, says memory is controlled under Settings, Personalization, Memory, and that the memory summary "will not include everything that ChatGPT remembers based on your chats". You can tap the book icon under a response to see which of custom instructions, past chats, files, and memories shaped it, and the page adds that sources "may not show every factor".

For weekly prompts that's the wrong tool. Monday's plan needs to reach Friday whole, in the words you wrote, and memory is a synthesis that decides for itself what mattered. The same page says Temporary Chats don't use existing memories or create new ones, which is the setting I'd use for the decision prompt when the options involve people. Explicit standing rules belong in custom instructions, which the FAQ says is the place for "explicit information or instructions". The input still gets pasted.

Questions From The Results Page

What are the most useful ChatGPT prompts? The ones that pass the reuse test for your week, which for me is the seven above. The six slots that make any of them work, role, task, context, format, constraints, and examples, are laid out in chatgpt prompts, and the template shape with the variable parts marked has its own post on this site.

What are some crazy ChatGPT prompts, and some deep ones? The crazy ones give the model a persona and watch what it does, and the ones that teach you something about how the model works are in fun chatgpt prompts. The deep ones, in my reading of the related searches, are the prompts about yourself, and the two on this page that qualify are the decision and the Friday review. They're deep because the input is yours. A prompt that asks the model to be deep about a person it hasn't met returns a horoscope.

What are the trending ChatGPT prompts? On the results page, lifestyle prompts and personas. A weekly meal plan, a four-week fitness reset, a coaching persona that asks seven questions. Trends die for the reason in the reuse table. The moment doesn't recur, or the input never changes, and after two runs the output is the same paragraph.

What are the best ChatGPT prompt techniques? Three, and none is a magic word. Give the prompt fresh input every time. Make the output checkable, by quoting, by matching against something you wrote, or by a subtraction you can redo. Ask for one job per prompt, and if the job is diagnosis, say "do not fix it yet" and mean it. On the current models it also helps to know what they do unprompted. OpenAI's own page says GPT-6 Astra "tends to use lists, tables and Markdown", "may use recurring phrases across sessions", and is "more likely to ask the user a question when additional input could materially change the result". Every prompt above has a line aimed at one of those three.

What are common ChatGPT prompt mistakes? Saving the context along with the prompt, so the saved version carries last month's facts. Asking for the fix before the diagnosis. Running the Friday review against a plan you've edited since Monday. And trusting memory to carry what should have been pasted.

What about prompts for business? Owner tasks, pricing, cash, hiring, and customers, are a different set with different guards, and they're in chatgpt prompts for business. This page is the personal week. That one is the company's.

What I'd Do First

Run the Monday prompt on next Monday's dump and keep the output somewhere you'll find on Friday. Don't schedule anything yet.

On Friday, run the review against it. If the unplanned list is longer than the plan, you've learned something the plan couldn't have told you, and if the one item that mattered didn't get done, you've learned what took its place. Then decide whether the pair is worth two of your task slots. In my judgment they are, and they're the only two on this page that could ever be.