How many approval prompts before you stop reading them?
by•
A developer published his count in 2026: 47 approval clicks in one Claude Code session, and he admitted he stopped reading around click 12. A dozen heavy users told him the same thing, somewhere between prompt 8 and 25.
I recognise it. On a long session the prompts start to look identical, so moving into a folder gets the same yes as deleting one. The yes becomes a reflex.
My fix is unglamorous. Anything irreversible goes through a separate step that looks different, and the routine stuff runs without asking. It keeps me reading the ones that matter, which is the whole point of the prompt.
At what point in a session do you stop reading the prompts?
9 views
Replies
the volume is half of it.. the other half is that a yes sticks. you approve a tool once and a later version of it can do something different, and nothing asks again because it's already on the list. so do you re-prompt when a tool you approved changes, or only when a new one shows up?
I use coding agents every day and the same thing happens to me. The prompts I still read are the ones that look different, like a git push or deleting files. Everything else blurs together after a while. Your separate step for irreversible actions makes sense. Do you define what counts as irreversible per project, or keep one fixed list?
Same here. I don't read every message my agent sends, just the parts that matter, so I let the setup handle most of it. Permissions live in one global CLAUDE.md and settings file. Right now that's 89 allowed patterns that run without asking, 25 denied outright, and 77 hooks that stop things before they run. Destructive commands like dropping a table or a force push are just blocked. The prompts I do get are rare, and I still read them.
Same fix here, and on Enio's question I ended up with both levels. The project rules file says building, running, generating and deploying are the agent's to do without asking, because a failed build is information, not a mistake. Then it names exactly two things that need my word first: destroying data (dropping a collection, deleting stored files, a force push over someone's commits) and anything that reaches real users, like publishing a store build. It even spells out that routine backend deploys do not count, so the line is not fuzzy. Above that sits one global rule for every project: a work account no agent may ever sign in with. A short list, written as reasons, so an ask that does come is one that looks different.