Thanks Wyndo and Nick, this is an excellent post. And it totally makes sense, why do we think that AI can work arrives for us without being onboarded properly. We wouldn't expect that of our human collaborators would we? 🙏
Exactly. The tricky difference is AI doesn't signal confusion the way humans do. A new hire asks clarifying questions. AI just... guesses with confidence. So you have to force the intake conversation or it skips straight to producing plausible garbage.
Love the team connection. Same principle applies: the questions that matter most are the ones that catch misalignment before someone spends three days on the wrong thing.
For AI, I prioritize assumption-surfacing questions. "What do you think I want here?" often reveals that AI is solving a slightly different problem than I intended. Better to catch that at minute one than draft three.
Thanks Nick, Wyndo. As always a great insightful article.
The "confident idiot" line made me laugh because it's so accurate.
I've been doing something similar with my n8n workflows. Before I let AI write any automation logic, I make it list out what it thinks the workflow needs to do. Saves me from rebuilding the whole thing later when I realize it assumed the wrong trigger or data format.
One thing that's helped me: I keep a doc with all my past corrections. Stuff like "don't use corporate speak" or "always start with the outcome, not the setup." Just paste that at the start of new chats or us Claude/ChatGPT projects and you skip some back-and-forth.
Same idea as your CORRECT step, just a bit external to the chat I think
The corrections doc is exactly right. That's the part most people skip because it feels like extra work. But it compounds. Every correction you document teaches AI your patterns for next time.
You're basically building a running preferences file. I call it a Voiceprint when it gets systematic enough. Same principle: externalize the corrections so you're not re-teaching the same things every session.
Absolutely.. it keeps improving, key thing is to keep recirculating the feedback into it and keep improving it as you learn even if it take extra 15 minutes.. as you said it will compound
Thanks Ilia! The prompts took embarrassingly long to get right. Turns out "ask questions first" is simple advice that's weirdly hard to prompt for. AI really wants to skip to showing off.
Love this practical Feed-Reflect-Correct framework — it really shifts the mindset from guessing to true collaboration with AI, which is something that’s been missing in most prompting advice. Thanks!
Thanks Robert! The reason it's missing from most prompting advice is that "slow down and ask questions" doesn't sell courses. "Magic prompt that 10Xs your output" does. The boring fundamentals don't have good marketing.
The "confident idiot" framing is perfect. The failure mode isn't that AI doesn't try - it's that it tries confidently in the wrong direction and you don't catch it until you've wasted 20 minutes.
Feed-Reflect-Correct maps well to how I think about it: give context, let it think out loud, then course-correct before execution. The reflection step is where most people skip straight to action and wonder why results are off.
Curious about the "Correct" phase - do you find it's better to correct with specific instructions ("change X to Y") or with outcome descriptions ("this should feel more conversational")?
Love this framework. I think so many of us are quick to paste and go... and that can spell doom in most cases. So important to reflect and correct along the way.
Actually I think many of us have always been lazy at communicating context and impatient at allowing questions from other humans. We have been able to get away with being sloppy in communication because of a mixture of reasons, e.g. we learnt to sense the non-verbal communication and read the room and the risk of getting fired/screamed at has us buckle up at learning the other person's communication to survive. But AI is not able to sense mood and communication nuances, so communicating with AI like we do with other humans exposes the gaps in our communication skills. My observation when I see other people's prompts, albeit occasionally, is they aren't even asking questions on how they can improve their prompts or even considering if their prompts are understandable by AI when there's total lack of punctuation in their sentences ❗❓
That's why I think the saying that has been thrown up a lot on the Internet "AI will not take your job but someone who uses AI (effectively) will" has some truth to it.
The advice "write better prompts" not only doesn't work, it's very difficult to apply in reality! It requires the person to be interested in improving this skill, regularly, and interested enough to develop their systems thinking skill/how to design the work on a systems thinking level. Then again, if one's livelihood is on the line, thus basic needs are on the line, that can be a giant motivator to learn new behaviours.
Yes. "Write better prompts" assumes you already know how to think in systems and communicate explicitly. It's like telling someone to "just draw better" without teaching them to see.
The livelihood pressure point is real too. Necessity forces skill development faster than curiosity. People learn explicit communication when vague stops working.
Context is king when it comes to using AI models. This framework is excellent! I have detailed explanations of my general goals loaded in a project and custom GPT to help with this.
It depends on the task. For nuanced tasks like writing "deeper" articles, creating automation logic, in-depth research, building a product, etc., I have AI interview me to get a more complete background of what I'm trying to do.
But some tasks don't need that much back and forth to get done. For most articles or shortform posts I write, I have enough of an idea of how it should turn out, and the AI has enough context about me to create good content.
Makes sense. The interview is highest-value when you don't fully know what you want yet. Writing a deep article, building something new, exploring a problem.
For tasks where the outcome is already clear in your head, the baseline context is usually enough.
Thanks Wyndo and Nick, this is an excellent post. And it totally makes sense, why do we think that AI can work arrives for us without being onboarded properly. We wouldn't expect that of our human collaborators would we? 🙏
Exactly. The tricky difference is AI doesn't signal confusion the way humans do. A new hire asks clarifying questions. AI just... guesses with confidence. So you have to force the intake conversation or it skips straight to producing plausible garbage.
I’ve seen the same principle at work in teams.
Taking a moment to pause, ask clarifying questions, and reflect cuts wasted effort and keeps outputs aligned with the real goal.
‘My Accountability Partner’ is all about creating that loop.
Curious, how do you decide which AI questions give the biggest ROI first?
Love the team connection. Same principle applies: the questions that matter most are the ones that catch misalignment before someone spends three days on the wrong thing.
For AI, I prioritize assumption-surfacing questions. "What do you think I want here?" often reveals that AI is solving a slightly different problem than I intended. Better to catch that at minute one than draft three.
That’s such a clean reframe.
Catching misalignment early saves more than time, it preserves momentum and trust.
Asking AI to reflect your intent first feels like the fastest way to make it a true thinking partner, not just a fast typist.
Thank you for reaching out Nick✨
"not just a fast typist"
I like that framing!
I’m glad you do Nick
Merry Christmas
Same to you! Happy Holidays!
🎉
Thanks Nick, Wyndo. As always a great insightful article.
The "confident idiot" line made me laugh because it's so accurate.
I've been doing something similar with my n8n workflows. Before I let AI write any automation logic, I make it list out what it thinks the workflow needs to do. Saves me from rebuilding the whole thing later when I realize it assumed the wrong trigger or data format.
One thing that's helped me: I keep a doc with all my past corrections. Stuff like "don't use corporate speak" or "always start with the outcome, not the setup." Just paste that at the start of new chats or us Claude/ChatGPT projects and you skip some back-and-forth.
Same idea as your CORRECT step, just a bit external to the chat I think
The corrections doc is exactly right. That's the part most people skip because it feels like extra work. But it compounds. Every correction you document teaches AI your patterns for next time.
You're basically building a running preferences file. I call it a Voiceprint when it gets systematic enough. Same principle: externalize the corrections so you're not re-teaching the same things every session.
Absolutely.. it keeps improving, key thing is to keep recirculating the feedback into it and keep improving it as you learn even if it take extra 15 minutes.. as you said it will compound
Amazing system that sounds like can actually work! And the prompts are spot on as well. Great guest post Nick!
Thanks Ilia! The prompts took embarrassingly long to get right. Turns out "ask questions first" is simple advice that's weirdly hard to prompt for. AI really wants to skip to showing off.
And thank you for reading Nick, appreciate your support!
Love this practical Feed-Reflect-Correct framework — it really shifts the mindset from guessing to true collaboration with AI, which is something that’s been missing in most prompting advice. Thanks!
Thanks Robert! The reason it's missing from most prompting advice is that "slow down and ask questions" doesn't sell courses. "Magic prompt that 10Xs your output" does. The boring fundamentals don't have good marketing.
The "confident idiot" framing is perfect. The failure mode isn't that AI doesn't try - it's that it tries confidently in the wrong direction and you don't catch it until you've wasted 20 minutes.
Feed-Reflect-Correct maps well to how I think about it: give context, let it think out loud, then course-correct before execution. The reflection step is where most people skip straight to action and wonder why results are off.
Curious about the "Correct" phase - do you find it's better to correct with specific instructions ("change X to Y") or with outcome descriptions ("this should feel more conversational")?
Depends on where you are in the process.
Early passes: outcome descriptions. Let it interpret "more conversational" and see what it tries.
Later passes: specific instructions. Once you know what's wrong, "change X to Y" is faster and cleaner.
The mistake is being specific too early. You close off options before you've seen them.
Love this framework. I think so many of us are quick to paste and go... and that can spell doom in most cases. So important to reflect and correct along the way.
Thanks for the good 😊
Actually I think many of us have always been lazy at communicating context and impatient at allowing questions from other humans. We have been able to get away with being sloppy in communication because of a mixture of reasons, e.g. we learnt to sense the non-verbal communication and read the room and the risk of getting fired/screamed at has us buckle up at learning the other person's communication to survive. But AI is not able to sense mood and communication nuances, so communicating with AI like we do with other humans exposes the gaps in our communication skills. My observation when I see other people's prompts, albeit occasionally, is they aren't even asking questions on how they can improve their prompts or even considering if their prompts are understandable by AI when there's total lack of punctuation in their sentences ❗❓
This is sharp. Humans learned to compensate for each other's communication gaps. AI hasn't.
It explains why "just write better prompts" advice feels frustrating.
The real skill being tested isn't prompting. It's explicit communication. And most of us have never had to be this precise before.
That's why I think the saying that has been thrown up a lot on the Internet "AI will not take your job but someone who uses AI (effectively) will" has some truth to it.
The advice "write better prompts" not only doesn't work, it's very difficult to apply in reality! It requires the person to be interested in improving this skill, regularly, and interested enough to develop their systems thinking skill/how to design the work on a systems thinking level. Then again, if one's livelihood is on the line, thus basic needs are on the line, that can be a giant motivator to learn new behaviours.
Yes. "Write better prompts" assumes you already know how to think in systems and communicate explicitly. It's like telling someone to "just draw better" without teaching them to see.
The livelihood pressure point is real too. Necessity forces skill development faster than curiosity. People learn explicit communication when vague stops working.
Thank you for this. I always research from other articles during my draft process. I always like to verify if the information I share is accurate.
Brilliant post guys. Thanks for sharing 🌟
Context is king when it comes to using AI models. This framework is excellent! I have detailed explanations of my general goals loaded in a project and custom GPT to help with this.
Smart setup. The project context handles the "who I am and what I care about" layer.
Curious: do you still have AI ask clarifying questions per task, or does the loaded context cover enough that you skip straight to generation?
I've found even with solid baseline context, task-specific questions catch assumptions I didn't anticipate.
It depends on the task. For nuanced tasks like writing "deeper" articles, creating automation logic, in-depth research, building a product, etc., I have AI interview me to get a more complete background of what I'm trying to do.
But some tasks don't need that much back and forth to get done. For most articles or shortform posts I write, I have enough of an idea of how it should turn out, and the AI has enough context about me to create good content.
Makes sense. The interview is highest-value when you don't fully know what you want yet. Writing a deep article, building something new, exploring a problem.
For tasks where the outcome is already clear in your head, the baseline context is usually enough.
The loop is a tool, not a ritual.