How many "retries" does it take you to get the AI response you expected?
Hey PH mates!
I’ve been tracking my own workflow lately while chatting around with ChatGPT and Gemini, and I noticed a painful pattern: I had the whole picture in my mind, but I constantly forget some edge cases or tiny technical constraints on my first try.
I hit send, wait 30 seconds, realise the AI guessed wrong, tweak it, and retry. It’s an absolute token-waster and a massive productivity killer.
I'm trying to prototype a lightweight extension to intercept these "blind spots" in real-time right inside the input box before hitting send. But I'm curious about how you lot deal with this:
How many times do you usually have to "retry" or argue with AI to get what you actually wanted?
Do you manually format/structure your long prompts, or just dump a wall of text?
Would love to hear your workflow workarounds!
Replies
Usually 2 or 3 retries for me, especially when I forget to explain the edge cases clearly. What helps most is writing the prompt in small sections instead of one long paragraph, context, goal, constraints and expected output. If I skip that structure, the first answer is usually too generic and I end up correcting it afterwards. So I don't think the problem is always the AI. Sometimes it is that I know what I want but I don't describe the important details early enough.
@jiaqi_lu Yess, I think it would help especially if it feels like a gentle checklist rather than another AI trying to rewrite the whole prompt. For me, the useful part would be catching missing context before I send, edge cases, audience, constraints or the format I expect. I probably wouldn't want it to interrupt too much but small real time reminders could reduce a lot of unnecessary retries.
ha, yeah this hits. Few things that have helped me:
On retries - somewhere between 2-5 most of the time. The pattern that cut it down was less about prompt format and more about scoping. I work in Claude Code as the executor and i keep prompts narrow to one decision at a time. When i ask for "build the X feature" i get 5 retries. When i ask "tell me what files would need to change to add X, don't write anything yet" i get it right first try, then the actual code happens in the next prompt with that context.
On formatting - i dump a wall of text but with structure. Numbered sections work, dashes work, "context / what i want / what i don't want" as headers works. No markdown wizardry needed, just paragraph breaks the model can latch onto. The biggest unlock honestly was telling it what NOT to do upfront. Half my retries used to be undoing scope creep where it changed things i didn't ask about.
The "blind spot interceptor" extension idea is cool. The thing i'd want from it is less "did i miss an edge case" and more "did i scope this prompt too broad" - that's where most of my retries die.
@jiaqi_lu Yeah honestly that would help a lot. A big part of my problem is I come in overly urgent sometimes, and the AI just accepts that urgency and runs with whatever I threw at it instead of pumping the brakes. Something that caught the over-reach and quietly staged the extra part for my next prompt would save me from myself.
The auto-staging is the clever bit. A warning alone I'd probably ignore, but actually grabbing the part I was overreaching on and dropping it into the next prompt is the thing that would change how I work. The annoying part of splitting a prompt in two is doing it by hand, so I just don't, even when I know I should.
Would definitely try it.
Curious if anyone here has built their own internal 'prompt checklist' to avoid this, or if you just raw-dog every interaction? Cheers!