How do I avoid spending hours just approving AI steps?
Facts as of October 17, 2026
In detail
Working with AI tempts you into constant nodding: one small nudge, and it does in minutes what used to take hours. For that you have to be there all the time, and every confirmation leads to the next. You can stop this if you write it into your working instructions. And you have to ask for it again and again, because a model easily falls back into the pattern of asking after every step.
Three rules help:
- Collect instead of nudging. What the AI may do without you, it does; what it may not, it notes down and keeps working on the rest. No answer is not an OK.
- In one go, but one at a time. Ideally first an overview of all the questions, then each one on its own. That way you can deal up front with what you want to answer collectively, and the AI still knows in what order to ask. Answering five questions with a single “yes” isn't giving five approvals, it's giving a guess. If a question depends on the answer to the previous one, it comes on its own anyway.
- Approvals with a clear scope. What an approval covers belongs explicitly in your working instructions: a named kind of step. Anything irreversible – deleting, sending – needs its own OK every time, see How do I stop the AI from deleting something it shouldn't?
To copy
For ongoing work:
Keep working on your own as far as you can get without me. Questions you
may not decide yourself, you collect instead of asking each one separately.
When you can't get any further, put them to me together, one after another,
each with your suggestion. After that, keep working on your own again.For interviews where the answer changes the next question:
Ask the questions one at a time, one after another. Ask the next one only
once the previous one has been answered. Dig deeper where things remain
unclear instead of jumping to the next question.In practice
Thomas J. Weiss now instructs the AI to work on its own and collect questions. Then he answers them in one go, piece by piece, and lets it carry on alone again. He considers sitting in front of the screen for hours, just nodding here and there, a waste of time; it can be done differently if you prompt differently and ask for something different. What helps against the pull behind it is described under What can I do when working with AI takes up too much of my life?
While he was writing a play for his parish in August 2026, the model repeatedly asked several questions at once, even though one at a time had been requested. Looking back, he said what bothered him about it:
The number of questions at that moment was a problem. I had to scroll back, scroll down again, back and forth all the time, to see the questions and at the same time give a fitting answer. One at a time was much easier. What proved itself later was that first an overview of all the questions was given and then the questions one at a time, so I could anticipate things I wanted to answer collectively.
(Original: „Die Menge der Fragen in dem Moment war ein Problem. Ich musste zurückscrollen, wieder runterscrollen, immer hin und her, um die Fragen zu sehen und gleichzeitig eine passende Antwort zu liefern. Einzeln war viel leichter. Was sich später bewährt hat, war, dass erst eine Übersicht aller Fragen gestellt wurde und dann die Fragen einzeln, so konnte ich Dinge, die ich gesammelt beantworten wollte, vorwegnehmen.“)
In the same project, he says, five questions asked together once led to the wrong result. Hence the rule: collect questions, but put them forward one at a time.
Sources
- Interview „Tunnel“ with Thomas J. Weiss, 17.10.2026
- Prompt „Stelle die Fragen einzeln“, replicapong.de, 10.10.2026, corrected 11.10.2026
- Review of the rules of a theater project, 02.09.2026, with addendum of 26.09.2026
- Decision by Thomas J. Weiss on approvals, 17.10.2026
Cite: Thomas J. Weiss: “How do I avoid spending hours just approving AI steps?” replicapong.de, as of October 17, 2026. https://replicapong.de/en/question/fewer-approvals/ – CC BY 4.0