When to stop prompting and do the task yourself

Prompting has a clock. When a fact must be true, or when the message needs your feelings, your relationship, or your signature, close the chat and do the task yourself. Nineteen tone drafts for twelve lines is the stop signal.

Imagine you are apologizing to a client about a wrong price. Your ChatGPT sidebar shows 19 messages in a thread titled “apology email tone.” You pasted the first try a little after four o’clock. Forty-seven minutes later the draft is still 12 lines and still wrong.

Prompting has a clock

A prompt is cheap in typing and expensive in minutes. You already had a usable intern brief with a task, an audience, facts, a format, and things to avoid. You also had a feeling that you could prompt your way into sounding like a person who was sorry. That feeling has no format line, so the box will keep offering synonyms until the workday ends.

Write two numbers before the first Send: the time on your screen, and the minutes the job would take you with the chat closed. In this story your pair is the start time and 6 minutes. Once the thread is older than 6 minutes, the model is no longer saving you anything, and it is billing you in attention. Free plans charge that way, and so do paid plans. Tokens, the units AI companies use to meter text, are a later problem covered in Practical AI. Today the unit is the clock on your screen.

The earlier post on iterating when the first answer is wrong taught you to keep going when the first answer misses. That loop is real, and it also has an exit. You get two useful turns: one brief and one named correction. “Cut the workshop, keep the $14,800, 12 lines, no partnership paragraph” is a named correction. “Make it more authentic” is not. If you cannot say the next edit in one sentence, you are arguing, and arguments belong in Gmail, where the send button carries your name.

Rule of thumb: If the argument with the model is longer than the job, close the chat and type the message yourself.

If you lost track of the clock, count turns instead. Three tone passes is a good place to close. Five “try again” messages is a good place to close. Nineteen messages for 12 lines is a close you already missed. The stop is simple arithmetic and has nothing to do with your personality.

Keep prompting, stop, or never start

Every job you bring to ChatGPT, Claude, Gemini, or Grok falls into one of three groups. A fourth move sits beside them, which is to leave the tab and open the file. The groups are about the work, and the fourth move is about truth.

Keep prompting when the facts are already in the paste and you still need a shape. That covers the study outline, the shopping grid, and the 80-word rewrite from the previous post. The intern test from the first post on talking to AI like a clear human still applies. A study outline that takes you 20 minutes alone can take 8 in the chat, so stay. An outline you have been poking for 40 minutes, because the model keeps adding a unit you did not assign, is a stop.

Stop when you already have something you could send and the new messages are about warmer, more authentic, less corporate, or more like me. Those adjectives have no target you can check, and the box will generate a new 12 lines forever. Paste the last true draft into Gmail, write the two sentences only you can write, and send it. Never start when the work is a feeling, a relationship, or a signature. The model can copy the shape of “sorry,” but it cannot be sorry, and your client can hear a template. If you need a first draft of a hard email, that is a keep with a short clock, and you still rewrite the line that says you are sorry. If the whole message is the apology, skip the tab.

When a fact must be true

OpenAI’s own help page on whether ChatGPT tells the truth is blunt. When this post was written in August 2026, it said to treat the reply as a first draft and not a final source, and to verify quotes, data, technical claims, and references. Anthropic’s usage policy and Google’s Gemini rules say the same thing in different words. The model will speak in a confident voice about $18,400, but confidence is not a receipt.

In this story the $18,400 started as a typo in your own first paste. The model kept it, then rounded up a “partnership credit” that did not exist, and then invented a workshop to explain the gap. Each turn looked like progress because the prose got smoother, while the number got worse. The signed statement of work (SOW), the contract that lists the agreed price, sat on your second monitor and never changed. The cheap move was to switch windows and look, which takes about forty seconds and shows $14,800. You do that after 48 minutes of tone editing.

Facts that must be true leave the chat. That includes money, such as prices, credits, invoices, and payroll, anything that will hit a card or a purchase order. It includes names of people, companies, product codes, and course codes. It includes dates and times, such as a freeze on Thursday, a send deadline, exam windows, and renewal dates. It also includes citations, meaning a paper, a clause, or a URL you will paste into someone else’s inbox. If you did not open the page, you do not have a source. The honesty and citations post in the safety series goes further on school and work, and the rule here is smaller: no link in the paste means no link in the send.

You can still draft around those facts. Paste the true number and label it “Facts, do not add any,” the way the first post taught. If the reply adds a second number, delete it instead of debating it. If you are unsure which number is true, the chat cannot settle it, but the SOW can, and so can the syllabus or the bank app. When the vendor’s own window offers a search result, open the link. Health, legal, and money decisions have a harder limit, described in AI safety for everyday life. You do not need that whole series to handle $14,800. You need the PDF.

Feeling, relationship, signature

Three kinds of job share a failure. The box can produce the right number of lines and the wrong person.

Feeling

Sorry, proud, scared, and grateful are states that live in your body and in the history you share with your reader. A model has a pile of apology emails from the public web, and it will give you “I value our partnership” because that phrase is common. You have never said “partnership” to this client. You say “the PO (purchase order, the document a buyer sends to approve a purchase)” and “the dock.” That fluent paragraph is a giveaway, and the client would read it as a template, which is worse than a blunt six-line correction.

Relationship

Think of a birthday message, a condolence note, a note to a sister, or a note to a person you manage. The reader already has a voice sample of you, and they will hear the seam. You can ask the box for gift ideas, or for three restaurant facts that you then check on the restaurant’s site. You cannot hand off the sentence that says you remembered. If you are staring at a blank Slack message because the feeling is hard, the blank is the work. Ten honest words beat 12 lines of workshop language.

Signature

A signature job is an offer letter, a purchase order, a performance review, or anything that will be forwarded with your name in the From field and no chat log attached. A first draft of a review can be a keep, with pasted notes and a short clock. The judgment line, such as “meets,” “misses,” or “I am putting you on a plan,” is yours. Clicking Send on a vendor quote you did not read is a signature move, even if you never drew an S. In this story your correction email is a signature, because your client’s finance team will file it and the workshop sentence would have become a fact in a later thread.

I think this is the part of prompting culture that still sells the wrong promise. The promise is that a long enough prompt will turn the box into you. A long prompt can make the box a sharper intern, but it cannot sit in your client’s standup, and it cannot be the person who typed $18,400. That person has to send the 12 lines.

A two-minute stop rule

Keep this rule under the intern skeleton from the first post, and fill in the brackets when you open the chat. Run the bottom half after every reply. Two minutes is the judgment window, not two more minutes of prompting.

STOP RULE (run after every reply)
Clock started: [HH:MM]
Solo time for this job: [N minutes]
Turns so far: [n]
Last reply still needs: [one sentence, or NONE]

If minutes in this thread > solo time: close chat. Type the artifact.
If a number, name, date, or citation must be true: leave chat. Open the source.
If this is a feeling, a relationship, or a signature: do not start, or stop now.

Two-minute pass:
1. Does the last reply contain a fact I cannot check in this window? Cut it.
2. Would I send this to [named person] in 60 seconds? If no, write it myself.
3. Am I asking for warmer / authentic / more like me? Close. That is a feeling job.
4. Named next edit in one sentence? If I cannot name it, I am arguing. Close.

This block forces the two numbers onto the page before the model can hypnotize you with a smoother paragraph. In the story you never wrote down solo time. You would have seen 6 against 47 while the workshop was still on screen. “Make it more authentic” fails the one-sentence line, while “Delete the workshop, keep $14,800, 12 lines to the client” passes, and then you are done in one turn. Put the block in the same note as the intern skeleton. It works in ChatGPT, Claude, Gemini, and Grok because it is a rule for you and not a hidden setting. None of those products will stop you at minute 7, so you have to.

A worked example: 47 minutes for 12 lines

Here is the same mess with two paths. One is what happened. The other is the stop rule together with the intern brief.

Here are the 12 lines you actually send, with the chat closed:

[Client contact],
I quoted $18,400 on the Q3 add-on this morning. That was my typo.
The signed SOW is $14,800. That is the number.
I am sorry. I should have checked the PDF before I hit send.
Corrected line is attached. Happy to jump on a 10-minute call if Holt needs a new PO today.
[Your name]

That is six sentences. It has the number, the sorry, and the attachment, with no workshop. Your client already knew the $14,800 and just wanted you to say it. The 47 minutes were spent trying to rent a feeling, and the 6 minutes were spent being the person who made the typo.

Here is a second case from the same afternoon, because you had one more tab open. It is your sister’s birthday, and ChatGPT has a 90-word note ready with a joke about “another trip around the sun.” Your sister hates that line, and you two have a running gag about a burned lasagna from 2019. That gag is a fact about your relationship that the box does not have and that you should not paste into a vendor’s chat. You close that tab in 20 seconds and send an 11-word Slack message. That is the never-was-a-prompt-job group working in under a minute, which is the point of naming it before you type.

Common mistakes

• Opening a chat for an apology, a condolence, or a birthday, then spending the evening on tone.
• Letting minutes pass without writing your solo time on the page. No clock means no stop.
• Treating “more authentic” as a named correction. It is a feeling with nothing to check.
• Negotiating a number with the model while the SOW is on the other monitor.
• Pasting a citation or a URL you have not opened, then forwarding it.
• Running the iteration loop with no exit. Iterate twice, then decide.
• Asking four vendors the same warmer prompt. That is four clocks and still one apology you have not written.
• Signing or sending a draft you have not read because the first sentence sounded like you.

How to practice this week

Pick one job you already did in a chat this week. Write the two numbers you skipped, which are the start time and your solo time, and count the turns. Put the job in the three-column table. If it was a keep, save the named correction that would have been turn two. If it was a stop, close the next similar thread at the first request for warmer. If it was a never, do the next one in Slack with the sidebar shut. Paste the stop rule under your intern skeleton. This series ends here, and the next post is chat history vs memory vs projects. The series index is Prompting for everyone, and the post before this one is Study, work email, plans, and shopping research. For more paths, see Learn and AI for writing and questions.

Quick recap

• Write your start time and solo time before the first Send.
• Allow two useful turns, then stop if the thread is older than the job.
• Keep prompting when the facts are in the paste and you still need a shape.
• Leave the chat to check money, names, dates, and citations. The SOW beats the sidebar.
• For feeling, relationship, and signature work, do it yourself.

Series notes

This is Part 5 of Prompting for everyone. Previous: Iterate when the first answer is wrong. Hub: Prompting for everyone.

Sources

Research and further reading used for this article:

• OpenAI Help: Does ChatGPT tell the truth? (first draft, not a final source; verify quotes, data, and references)
• OpenAI usage policies (what the product is for, and the limits that stay on you)
• Anthropic usage policy (same shape on Claude: you own the send)
• Anthropic: Getting started with Claude (clear asks, then your review)
• OpenAI Help Center and Google Gemini (product behavior moves; confirm the week you write)
• NIST AI Risk Management Framework (judgment and validity sit with the person who uses the output)
• AMS: Prompting for everyone (this series; Part 5 is the exit)
• AMS: Iterate when the first answer is wrong (the loop this part closes)
• AMS: Memory, projects, and files (the next series after this one)

Merch on Etsy

Open store →

Naive Bayes Coffee Mug View on Etsy → Standard Deviation White Ceramic Mug View on Etsy → Mean Squared Error Shirt View on Etsy → Git Version Control Shirt View on Etsy →