Ask AI for one thing at a time

I remember few months back I asked ChatGPT for a PDF. It made the PDF. Then I checked it. The formatting broke in the third paragraph. The content was dull.

So I blamed my prompt. I rewrote it. I regenerated. I regenerated again. At 17 tries ChatGPT cut me off for the hour.

Then I found the video that sold me the idea. 275,000 views, one message in, a finished deck out. So I read the comments. One person said it works, but not at the level shown. Another said the result was plain, thin on information, no pictures.

Nobody in that thread was blaming their own prompt. I was the only one blaming myself.

Here is what I got wrong. I kept fixing the prompt, and the prompt was not the problem.

One message was carrying two jobs. Write the thing, and build the file. Those are different jobs, and each one fails in its own way.

So now I work in stages. One clear job per stage. Finish it, check it, then start the next one.

Each stage carries one set of instructions, so less gets dropped. And when something goes wrong, I know which stage did it. That second part is what actually changed my work.

Short on time? Get the one-page checklist. It shows where to split a request, when to stay in the same chat, and when to start a new one.

So I tested it properly.

I put our last five polls into one file. 1,664 votes and 56 of your comments. Same file, same model, memory off, fresh chat both times.

Four jobs: analyse the feedback, recommend a direction, make a chart, draft the announcement email. One run got all four in a single message. The other got them one at a time.

I expected the single message to get numbers wrong. It did not. I checked every figure by hand against the source, and both runs were clean.

The difference was somewhere I was not watching. The one-message run wrote five bullet points of analysis. The staged run wrote five full sections.

And the one-message run dropped a poll. Poll 3 had 624 votes, more than a third of the data. It never appeared. It was not called weak or off-topic. It was just not there.

Nothing was wrong. There was just less of it, in the exact place where the thinking was supposed to happen.

I ran the same comparison on four more jobs. Same direction every time.

One model, one run each side, and I graded it myself. That is why I wrote my expected answers down before I started, in a file I did not open until grading.

Here is the shape I use now. Same four jobs, four messages.

Message 1. The thinking, alone. Analyse this. Nothing else.

Then I read it before I ask for anything. I am not proofreading. I am looking for what is missing, not for what is wrong.

Message 2. The judgment. Based on what you just found, what should we do.

Message 3. The chart.

Message 4. The email.

Same chat all the way through, because every step serves one outcome. A new subject or a new client gets a new chat.

Two lines do most of the work. End each message with "only this, nothing else", and when a stage turns your text into a file, chart, or image, add "use the text above exactly, do not shorten or rewrite it".

If you already run this inside a Project or a custom GPT, put that second line in the system prompt. Just know that an automatic chain removes the read step between stages, and the read step is where I caught the missing poll.

Total cost: three extra messages.

Do not stage everything. Most of my messages are still single messages.

One message is enough when the thinking is already done and you are pasting it in. When the output is short, like one email or one caption. When you gave a template and only want it filled. When the format is plain text. And when nobody but you will see it.

The moment a request contains the word "and", read it again. That is usually where the second job is hiding.

Staging is not a repair kit. Here is what it does not touch.

It does not fix a wrong file type. If you need PowerPoint and the tool makes HTML, no prompt order changes that.

It does not add missing features. Your own images, animations, and brand fonts are tool limits, not prompt limits.

And it will not save you from a wrong number. Staging gives you a place to catch one. You still have to look.

Take the request you were about to send. Find the word "and" and cut it there. Send the first half alone.

Read what comes back. Look for what is missing, not for what is wrong. Then send the rest.

Five minutes, one task. That is the whole test.

Keep this beside ChatGPT or Claude. Print it for your desk, or save the image to your phone.

Know someone who is on try number 12 right now? Forward this to them.

One more thing from the test. My real mistake was not the stacked prompt. It was that I graded the output by reading it, and reading only catches what is wrong, not what is absent.

Hit reply and tell me one task you have been stacking into a single message. I read every reply.

Dan Rice · AI Signal Read once. Use AI better all week.