I can write from zero? I do many thing form zero in my past experience. But the speed up is incredible. My actual project multiplyer is around 100 x. For example I cannot able to implement |> into TS ( based Go ). So my short answer is no.
AI testing storyteller. Writing the 36 Stratagems series — ancient Chinese military tactics meet modern AI system meltdowns. 16yr QA → building AI test framework.
This got me thinking🤔. I wonder if there's a difference between reviewing AI's work and actually thinking through it.
If we just accept whatever AI gives us, then yeah, our problem-solving skills might get rusty. But sometimes reviewing means questioning a design choice, trying an edge case nobody thought of, or figuring out why the AI got something wrong.
I've been working with AI-driven testing, and one thing I've learned is that getting AI to do the work is only half the job. You still need to know how to challenge what it gives you.
Maybe the real question isn't whether we can still build everything from scratch. It's whether we can still recognize when AI gets it wrong.😆
AI testing storyteller. Writing the 36 Stratagems series — ancient Chinese military tactics meet modern AI system meltdowns. 16yr QA → building AI test framework.
Haha, that's a good point! You need to have made a few mistakes yourself before you can spot them in someone else's work. Maybe that's why experience still matters, even when AI can do so much of the actual work. 😅
Yes. Every time you make a mistake, you learn what to watch out for. This model has already looked at a million architectures, but has never had to correct a bad design for itself at 2am 😄
I think there is an even scarier version of the question for people who are new to coding. It’s not “Could I still do this from scratch?” It’s “Did I ever learn to do this from scratch?” When AI is there from day one, it’s easy to skip the struggle and not even notice.
Honestly, with AI around, doing anything from scratch is going to feel harder and much slower for most devs. At this point, I think writing every line yourself matters less than having strong foundations. You need to understand the code, even if you didn’t write all of it. That’s what saves you when the AI gets it wrong.
That’s the creepier one, you’re right. “Did I ever learn it?” never makes you aware of the missing piece.
And yeah, the understanding of foundations is what’s missing. You don’t need to write every word but you do need to know what’s going on so that if the AI ever gets it wrong, you can fix it. Thanks Maame!
The METR finding is the one that stuck with me — experienced developers were 19% slower with AI tools while expecting to be 24% faster. It makes the reviewing-versus-building split feel less like a preference and more like a trap: the review feels like work, but it is the building that keeps the judgment sharp. Do you have a way of checking in with yourself on this, like deliberately doing a small task unaided now and then, or is it more of a gut feeling when the third prompt doesn't land?
That was a good observation, the METR numbers showed developers estimating that it would be 24% faster but actually ended up 19% slower.
In response to your question, it's mainly a gut instinct. When the third prompt doesn't resonate, that's my cue to save time and switch over to a pre-defined, smaller action that doesn't use AI, rather than continuing to be too dependent on the AI signal.
That third-prompt cue is a neat rule — it turns a vague feeling into a concrete stop condition, and switching to a smaller pre-defined action instead of re-prompting keeps you in the driving seat. It also pairs well with Kryeo's point in this thread about writing your own decision down first, so the gut check has something to compare against. When you switch to that smaller action, do you usually break the original task down further, or reach for a specific non-AI fallback like reading the code or docs directly?
That sounds like a good middle ground — the cue doesn't have to be a hard rule to be useful, just a nudge to switch from prompting to reading. Checking the code or docs before the next prompt is the part that usually pays off fastest in my experience. Have you noticed whether it changes the kind of prompt you write after looking at the code, or is it mostly about catching a wrong assumption earlier?
ormer IB/AP science teacher building tentus solo with AI tools. I write about how it works, what breaks, and what 18 years in the classroom taught me about memory.
Location
Remote
Education
ITT tech BsC
Work
Founder at tentus · Former IB/AP Biology & Chemistry teacher (18 yrs)
The fear makes sense, but I'd draw the line in a slightly different place. The skill doesn't fade because you use the model. It fades when you stop producing your own answer before you see one.
In teaching, that's a well-studied effect: students who attempt a problem first, even badly, remember the worked solution much better than students who read the solution straight away. The attempt is what makes the answer stick. Reviewing AI code has the same structure. If you read it cold, you're the student reading the answer key. If you spend two minutes first writing what you expect (the signature, the edge cases, and where you think it'll break), the review becomes practice, and the places where the model surprised you show exactly what you didn't know.
That's cheaper than a weekly "write code without AI" ritual, and it fits into normal work. Has anyone tried something like this on a team, and did it survive a deadline week?
ormer IB/AP science teacher building tentus solo with AI tools. I write about how it works, what breaks, and what 18 years in the classroom taught me about memory.
Location
Remote
Education
ITT tech BsC
Work
Founder at tentus · Former IB/AP Biology & Chemistry teacher (18 yrs)
Thanks, Daniel. I'm building solo too, so I haven't tested it on a team either.
For crunch weeks there's a shorter version: before you open the diff, write one line saying where you expect it to break. It takes about 20 seconds and still counts as producing your own answer first. Memory research calls this the generation effect: what you produce yourself sticks better than what you only read. Wrong guesses help too, as long as you check them against the real answer straight after.
If you try it, I'd like to hear where your misses end up clustering. That's usually the part of the codebase you understand least.
Thanks Juan. I appreciate the 20 second version. There are some weeks I can pull that off...and other weeks...not so much. I will give it a try and let you know if I notice any mis-estimation trends.
Hello! I'm Rasheed, a .NET enthusiast who's constantly looking for new things to learn and forget, that's why I started writing in the first place, but I love writing tho :-)
The problem with AI that is hammering the skill degradation more quickly is the speed at which devs are expected to deliver nowadays, like previous features used to be planned months in advance, then given spacious times or at least good segments of time on roadmaps or release schedules, there will be close deadlines on which you will be grinding day and night at your keyboard, you will introduce tech debt, bugs, bad code, but you are still in control because you know each one since you had already been in the trenches with these problems. However, now because the AI can whip up a feature in an afternoon like you mentioned, you are expected to be delivering more.
Larger features, bigger expectations, things are just expected to happen much more quickly because you simply have a super agent running to your side, so for your own self, you start reaching out to the easy way out, instead of spending hours and days wrestling a problem and potentially extending the already shrinking deadlines, you can have it done for you at a much quicker pace, simply shaving away at the question of "Can I do this myself with that much time?", we all just answer that with No simply and then move on to continue prompting and shipping.
Founder of Aladdin AI, a tool that finds jobs that fit you, writes tailored resumes & cover letters and applies for you. 100 free applications a month, no paywall, no credit card. Building openly.
I build Aladdin, a job agent. Your debugging point is the one I keep coming back to. The agent reads the actual form for each application, and when it cannot read a field it asks the user instead of guessing. That only works if someone still knows what a correct answer looks like. Reviewing does build one skill, spotting a wrong field or a made-up detail. It does not build the other one, knowing what to do when the agent is stuck. Your METR point fits that. The model is fast on the familiar part and slow where judgment lives.
Founder of Aladdin AI, a tool that finds jobs that fit you, writes tailored resumes & cover letters and applies for you. 100 free applications a month, no paywall, no credit card. Building openly.
That is the exact distinction. Review catches a wrong answer, not a worker that quietly stopped. In Aladdin the fresh worker reads the actual form, so when it cannot answer a field it asks the user instead of guessing. The question itself is what makes the stuck case visible.
I research AI tools, APIs & agents — comparing specs, pricing and real user reports — then publish honest rankings so you don't waste money. Research-first, never sponsored fluff.
The third or fourth prompt is exactly where it hits me too — fifteen minutes of thinking replaced by fifteen seconds of accepting, and I can't even remember what I was deciding. I don't miss boilerplate, I miss the fumbling. Lately I've been doing the first sketch on paper before I open the agent, just to keep that muscle. Does that actually help, or am I just slowing myself down on purpose?
"I miss the fumbling" is probably the best comment in this whole thread, I believe that drawing out the plan on paper helps to force the user to make a choice before the bot does, and this is what gets lost when it gets automated.
It does slow down the process but only in the moment of thought which seems like a fair trade to me.
hate to break it to ya but AI isn't actually better
if AI makes one bug, tell it to fix the bug, it will start to ruin your program. the more you say "fix this" the more wrecking balls it gets to throw at your project
it may not make mistakes very oftem but once it makes one there's no way out.
excuse me how is writing code without AI something uncool?
I think your "wrecking ball" point is spot on. Each time you have to "fix" something you layer more stuff over that's actually moving the code further from the original.
In terms of your second point, I didn't mean that it's not cool, I meant it's not glamorous. I do think that it's a valuable skill to write code yourself - more so than it's ever been.
Two founders who write the code. We build headless commerce, CMS platforms and web apps on Medusa, Directus and Next.js. You own the repo, the data and the infra.
"A model that has seen a million architectures but never lived inside one" is the line that stuck with me.
What's helped us is being strict about where the thinking happens. Generating code is fine, but the decision gets written down by a human first: what we're choosing, what we rejected, and why. Before anyone opens a prompt. If I can't write that paragraph, I don't understand the problem yet, and no amount of accepted suggestions fixes that.
Debugging is the other one worth protecting. Reading the stack trace and forming your own guess before asking the model feels slow, but it's the closest thing left to that old fumbling-in-the-dark part.
Do you have a rule like that yourself, or is it more of a feeling you keep checking in on?
One small habit could be writing down your own approach, even three lines, before prompting, then comparing it with the model's answer. That gives you a chance to spot gaps in your reasoning before a plausible solution makes them harder to notice. Could that work as a team learning habit, rather than only solo practice?
The performance-review cover is the part that makes the fade hard to notice. Shipping still looks fine from the outside, and the empty shelf only shows up when a weird bug or a novel constraint appears. A scheduled unaided check would catch that earlier than a failed third prompt: pick one recent incident, even one already fixed, and walk the cause from the log without opening a chat. If a hypothesis does not form, that is the signal. Would a calendar reminder work better than the gut feel you described, or does the gut check already come often enough?
The reviewing vs building split is the part that stings. Reviewing a 90% right answer feels like work but you skip the fumbling where understanding forms. The skill I'd worry about losing fastest is debugging when the model gets stuck, because that's when you find out what's still under the hood. Curious if you think reviewing compounds into anything over time or just plateaus
Spot on about debugging, that's when you actually find out what you know.
i think reviewing compounds for spotting bugs and bad patterns, kind of like reviewing PRs all day. it just plateaus when you have to solve something from scratch. reading a solution doesn't teach you the same way getting stuck does.
I can write from zero? I do many thing form zero in my past experience. But the speed up is incredible. My actual project multiplyer is around 100 x. For example I cannot able to implement
|>into TS ( based Go ). So my short answer is no.Thanks Peter. I like the honest answer. I like that some of the work you couldn't do on your own was opened up to you by the tool.
I am just concerned with what happens when the tool gets stuck - do you read the result or do you tell the tool to fix itself?
This got me thinking🤔. I wonder if there's a difference between reviewing AI's work and actually thinking through it.
If we just accept whatever AI gives us, then yeah, our problem-solving skills might get rusty. But sometimes reviewing means questioning a design choice, trying an edge case nobody thought of, or figuring out why the AI got something wrong.
I've been working with AI-driven testing, and one thing I've learned is that getting AI to do the work is only half the job. You still need to know how to challenge what it gives you.
Maybe the real question isn't whether we can still build everything from scratch. It's whether we can still recognize when AI gets it wrong.😆
fair point, scrutinising the output is half the job. building from scratch is how you learn what "wrong" looks like in the first place....😆
Haha, that's a good point! You need to have made a few mistakes yourself before you can spot them in someone else's work. Maybe that's why experience still matters, even when AI can do so much of the actual work. 😅
Yes. Every time you make a mistake, you learn what to watch out for. This model has already looked at a million architectures, but has never had to correct a bad design for itself at 2am 😄
I think there is an even scarier version of the question for people who are new to coding. It’s not “Could I still do this from scratch?” It’s “Did I ever learn to do this from scratch?” When AI is there from day one, it’s easy to skip the struggle and not even notice.
Honestly, with AI around, doing anything from scratch is going to feel harder and much slower for most devs. At this point, I think writing every line yourself matters less than having strong foundations. You need to understand the code, even if you didn’t write all of it. That’s what saves you when the AI gets it wrong.
That’s the creepier one, you’re right. “Did I ever learn it?” never makes you aware of the missing piece.
And yeah, the understanding of foundations is what’s missing. You don’t need to write every word but you do need to know what’s going on so that if the AI ever gets it wrong, you can fix it. Thanks Maame!
The METR finding is the one that stuck with me — experienced developers were 19% slower with AI tools while expecting to be 24% faster. It makes the reviewing-versus-building split feel less like a preference and more like a trap: the review feels like work, but it is the building that keeps the judgment sharp. Do you have a way of checking in with yourself on this, like deliberately doing a small task unaided now and then, or is it more of a gut feeling when the third prompt doesn't land?
That was a good observation, the METR numbers showed developers estimating that it would be 24% faster but actually ended up 19% slower.
In response to your question, it's mainly a gut instinct. When the third prompt doesn't resonate, that's my cue to save time and switch over to a pre-defined, smaller action that doesn't use AI, rather than continuing to be too dependent on the AI signal.
That third-prompt cue is a neat rule — it turns a vague feeling into a concrete stop condition, and switching to a smaller pre-defined action instead of re-prompting keeps you in the driving seat. It also pairs well with Kryeo's point in this thread about writing your own decision down first, so the gut check has something to compare against. When you switch to that smaller action, do you usually break the original task down further, or reach for a specific non-AI fallback like reading the code or docs directly?
Thanks. It's not quite a habit yet, more of an idea.
If I got to the third prompt then I'd like to check out the code/docs and try prompting again.
Kryeo's rule of 3 provides a benchmark for this.
That sounds like a good middle ground — the cue doesn't have to be a hard rule to be useful, just a nudge to switch from prompting to reading. Checking the code or docs before the next prompt is the part that usually pays off fastest in my experience. Have you noticed whether it changes the kind of prompt you write after looking at the code, or is it mostly about catching a wrong assumption earlier?
The fear makes sense, but I'd draw the line in a slightly different place. The skill doesn't fade because you use the model. It fades when you stop producing your own answer before you see one.
In teaching, that's a well-studied effect: students who attempt a problem first, even badly, remember the worked solution much better than students who read the solution straight away. The attempt is what makes the answer stick. Reviewing AI code has the same structure. If you read it cold, you're the student reading the answer key. If you spend two minutes first writing what you expect (the signature, the edge cases, and where you think it'll break), the review becomes practice, and the places where the model surprised you show exactly what you didn't know.
That's cheaper than a weekly "write code without AI" ritual, and it fits into normal work. Has anyone tried something like this on a team, and did it survive a deadline week?
This is better than what I wrote. More accurately: "The skill atrophies if you don't generate a solution first."
The 2-minute writeup of a probable solution is a low-cost, non-daily way to practice.
As a solo developer, I can't say how it fares during a crunch week with a team. I'll give it a try the next time I'm reviewing.
Thanks, Daniel. I'm building solo too, so I haven't tested it on a team either.
For crunch weeks there's a shorter version: before you open the diff, write one line saying where you expect it to break. It takes about 20 seconds and still counts as producing your own answer first. Memory research calls this the generation effect: what you produce yourself sticks better than what you only read. Wrong guesses help too, as long as you check them against the real answer straight after.
If you try it, I'd like to hear where your misses end up clustering. That's usually the part of the codebase you understand least.
Thanks Juan. I appreciate the 20 second version. There are some weeks I can pull that off...and other weeks...not so much. I will give it a try and let you know if I notice any mis-estimation trends.
Good luck with your solo build as well.
let's start a public repo where we rebuild Netflix by hand, haha.
will take us back home
I'm in! 😄😆. Let's start by building the login page.
By the time we're done with the video player, Netflix will have rebuilt itself twice.
Maybe let's try a simpler objective first. a to-do list app without any AI. That would also take us home as well. 🤣🤣🤣
The problem with AI that is hammering the skill degradation more quickly is the speed at which devs are expected to deliver nowadays, like previous features used to be planned months in advance, then given spacious times or at least good segments of time on roadmaps or release schedules, there will be close deadlines on which you will be grinding day and night at your keyboard, you will introduce tech debt, bugs, bad code, but you are still in control because you know each one since you had already been in the trenches with these problems. However, now because the AI can whip up a feature in an afternoon like you mentioned, you are expected to be delivering more.
Larger features, bigger expectations, things are just expected to happen much more quickly because you simply have a super agent running to your side, so for your own self, you start reaching out to the easy way out, instead of spending hours and days wrestling a problem and potentially extending the already shrinking deadlines, you can have it done for you at a much quicker pace, simply shaving away at the question of "Can I do this myself with that much time?", we all just answer that with No simply and then move on to continue prompting and shipping.
I build Aladdin, a job agent. Your debugging point is the one I keep coming back to. The agent reads the actual form for each application, and when it cannot read a field it asks the user instead of guessing. That only works if someone still knows what a correct answer looks like. Reviewing does build one skill, spotting a wrong field or a made-up detail. It does not build the other one, knowing what to do when the agent is stuck. Your METR point fits that. The model is fast on the familiar part and slow where judgment lives.
That's a very explicit way of saying it.
Being able to review helps with recognizing when the result is wrong but not when the agent gets stuck.
If your agent tries to ask the user rather than guessing, it makes the "stuck" case explicit.
Good example, thanks!
That is the exact distinction. Review catches a wrong answer, not a worker that quietly stopped. In Aladdin the fresh worker reads the actual form, so when it cannot answer a field it asks the user instead of guessing. The question itself is what makes the stuck case visible.
The third or fourth prompt is exactly where it hits me too — fifteen minutes of thinking replaced by fifteen seconds of accepting, and I can't even remember what I was deciding. I don't miss boilerplate, I miss the fumbling. Lately I've been doing the first sketch on paper before I open the agent, just to keep that muscle. Does that actually help, or am I just slowing myself down on purpose?
"I miss the fumbling" is probably the best comment in this whole thread, I believe that drawing out the plan on paper helps to force the user to make a choice before the bot does, and this is what gets lost when it gets automated.
It does slow down the process but only in the moment of thought which seems like a fair trade to me.
hate to break it to ya but AI isn't actually better
if AI makes one bug, tell it to fix the bug, it will start to ruin your program. the more you say "fix this" the more wrecking balls it gets to throw at your project
it may not make mistakes very oftem but once it makes one there's no way out.
excuse me how is writing code without AI something uncool?
I think your "wrecking ball" point is spot on. Each time you have to "fix" something you layer more stuff over that's actually moving the code further from the original.
In terms of your second point, I didn't mean that it's not cool, I meant it's not glamorous. I do think that it's a valuable skill to write code yourself - more so than it's ever been.
"A model that has seen a million architectures but never lived inside one" is the line that stuck with me.
What's helped us is being strict about where the thinking happens. Generating code is fine, but the decision gets written down by a human first: what we're choosing, what we rejected, and why. Before anyone opens a prompt. If I can't write that paragraph, I don't understand the problem yet, and no amount of accepted suggestions fixes that.
Debugging is the other one worth protecting. Reading the stack trace and forming your own guess before asking the model feels slow, but it's the closest thing left to that old fumbling-in-the-dark part.
Do you have a rule like that yourself, or is it more of a feeling you keep checking in on?
Actually, your rule is much better than mine - having to write down your answer before you have it is a much stronger constraint.
That one sentence is very powerful: "If I can't write that paragraph, I don't understand the problem yet"
Answering the question: It is mostly intuition based for now. But I am always verifying it and your rule makes it much more actionable.
One small habit could be writing down your own approach, even three lines, before prompting, then comparing it with the model's answer. That gives you a chance to spot gaps in your reasoning before a plausible solution makes them harder to notice. Could that work as a team learning habit, rather than only solo practice?
Yes, I think it can be used by a team. Let them write 3 lines of code first and compare it with the model output.
Review can be done by showing the 3 lines of the author, as well as the final version.
It will make it easier for the team to understand the changes, rather than showing only the diff.
The performance-review cover is the part that makes the fade hard to notice. Shipping still looks fine from the outside, and the empty shelf only shows up when a weird bug or a novel constraint appears. A scheduled unaided check would catch that earlier than a failed third prompt: pick one recent incident, even one already fixed, and walk the cause from the log without opening a chat. If a hypothesis does not form, that is the signal. Would a calendar reminder work better than the gut feel you described, or does the gut check already come often enough?
The reviewing vs building split is the part that stings. Reviewing a 90% right answer feels like work but you skip the fumbling where understanding forms. The skill I'd worry about losing fastest is debugging when the model gets stuck, because that's when you find out what's still under the hood. Curious if you think reviewing compounds into anything over time or just plateaus
Spot on about debugging, that's when you actually find out what you know.
i think reviewing compounds for spotting bugs and bad patterns, kind of like reviewing PRs all day. it just plateaus when you have to solve something from scratch. reading a solution doesn't teach you the same way getting stuck does.
Writing down my own approach before prompting is the habit I'll actually keep.