LeftRightThere
35146
764
41
I still code projects by hand. I use AI to supplement my tasks, but I’m deliberate with what it does. I coded software to mitigate vulnerabilities for like 15 years and I’ll be damned if I let a juiced autocorrect “decide” what my code does. I won’t be the one adding tech debt to the stack.
I think this tag was vibe coded
override367
The skills to be a good AI coder and the skills to be a good coder are completely different. If you write good tests for the AI that i must pass and write incredible design documentation, it is *dramatically* more effective than otherwise. Like night and day. You can reduce hallucinations to effectively zero if you're good at writing tests and docs
Oatmealman1
Any company that relies on AI deserves to crash
cousteau
AI has its uses. Basically all the stuff humans can't do or would be very hard for humans to do, but AI is good at and even better than humans. Coding is not one of those things.
Iwouldbenick
As an infrastructure engineer I hand write my network configs, always have. I only use AI to assist with powershell scripts that don't run. I am not very good with powershell commands. But I am very good at knowing when and how to use them in intune remediations over group policy settings.
RthereKittynadosN2020
I love RDJ!
JohnSmithterms
As a software dev with 20 years experience i approve this post https://media1.giphy.com/media/v1.Y2lkPWE1NzM3M2U1MWI3MGFiaWZ3enFha3FtMHlrZGY4Mm9odGxqdTM4enR1NzRjOTg1NiZlcD12MV9naWZzX3NlYXJjaCZjdD1n/diUKszNTUghVe/200w.webp
mygodhasabiggerdick
I took a bootcamp for Data Analytics. If I wasnt able to check my Python with an AI I dont think I would have found my errors. It did help me learn a bit, but people who code ONLY with AI... The haven't learned fuck all.
FreePalestineAndTheWorldFromIsrael
Telling someone else to do your job doesnt give you ownership over their work.
Telling something else to do your job doesnt give you ownership over the work.
quade
Downvote me for being out of line, but if you're not using AI to write software professionally, you'll be left behind and without a job.
TheSpectralPenguin
Fuck you AI shill.
LeftRightThere
I do use it, it’s part of my Lead AI/ML Engineer title. It’s great for assisting in code analysis, creating functions, fully formed README files. It can write entire apps, but AI isn’t trained on the best code, it’s trained on all the code available. Most code is garbage. AI is also terrible at the nuances of designing architecture.
If you want AI to write your code, you need to divert resources to VERY detailed requirements and DEEP dives into code review. The amount of work doesn’t change.
0xDEC0DE
skunkworkz
As an owner, and 25+ year developer, that's not happening here.
We aren't using this shit, and aren't going to use it (it's a terminable offense to feed any of our code or customer data to these companies).
Beyond the loss of critical thinking (well documented in lots of places), and serious issues with training the next generation of developers, the main issue is that there's no ethical way to use this stuff, so we won't use it.
zagibu
Unfortunately true. I'm already looking for a different profession, because I won't ever be a vibe coder.
UnicornsToxicRainbowFart
https://www.businessinsider.com/ibm-ceo-big-tech-ai-capex-data-center-spending-2025-12?op=1
LeftRightThere
Yeah, it’s totally a bubble that everyone is desperately trying to get rich on until it pops. The internet is becoming/has become a bot highway where humans aren’t the drivers.
https://en.wikipedia.org/wiki/Dead_Internet_theory
Velv3tThunder
AI is good for "how do I do this specific thing in this language/framework?" Not for building the whole fucking app.
LeftRightThere
And for providing options. Need a function or class, ask AI to make multiple different implementations to compare performance / maintainability.
SirenSinger17
I recently had a dude tell me he wrote 6 books with AI and he sincerely expected me to be impressed.
Noxnarie
I use AI to do repeating jobs like some arrays and JSON blocks. And I find AI to be best teacher for me because I can always question it's answers and it never gets tired of my stupid questions. Can't learn anything in school but arguing with AI is effective.
0xDEC0DE
I'd like to think I've had some successes with using AI as a teacher, but that's only if you heavily tune your agent preferences such that the default mode of "Dr. Flattery, the Always-Wrong Robot" is not active. Insist that it cite sources, push back when confronted, etc. otherwise it just backs down and agrees with you. That's what it's built to do, after all.
Noxnarie
Yeah that is most annoying thing about it but I've used (ChatGPT mostly) so long that it remembers tons of stuff I've asked it to remember so once the problematic code or function is finally working as it should, I ask gpt to remember our argumentative discussion about the problem and next time same problem or similar comes up, it performs much better with providing correct answer and explaining why it gave me that answer.
TheoryOfRelativeTea
I'm similar. The only reason I have ever used any 'AI' is to run down why I'm getting repeated traceback errors instead of spending 3 hours on stackoverflow and reddit looking for some hyper-specific answer. And I only submit what is absolutely necessary to get an answer. I also change the names of variables and functions. I don't want or need it to 'create' code for me, just aggregate all the solutions for my problem and present the most likely fixes for my symptoms.
0xDEC0DE
One of the more frustrating things for me over the years has been having to map "toy problem" example code into real-world use. With an LLM, those days are long, long, gone. Now you can give it your code, ask about what you want to do, and get a working example tailored against YOUR code, with YOUR algorithms, YOUR variable names, etc. Having the clanker contextualize it for you makes understanding SO MUCH easier. Try it!
skunkworkz
Never been happier to own my product, and my own company.
We aren't using AI in our development at all, and won't in the near future (really we're dumping big tech anywhere possible).
There are just too many ethical issues with AI, the data it was trained on, the datacenters, environmental issues, and the owners / founders are all bat shit crazy enablers of the current US regime.
AnimalRescueGuy
I don’t use AI in my daily life. People complaining about electricity costs and water scarcity, then turn around and ask Claude what they can do about it. smh
LeftRightThere
I have a personal AI server in my house that I do development on. I built a standalone solar energy system to power it. It has 2 days of standby power before it needs to pull from the grid. I learned so much about creating off grid solar because of this.
mutedia
I curse at claude a lot but I'm not going back to coding everything by hand. I build a plan and fill out the architecture and establish some guidelines and then I bounce off of claude. I still approve every modification and push back a lot as half the time there's something that doesn't make sense.
neverpostsoriginalcontent
@OP, I'm a full-stack dev but been working primarily in JS for ages, and I happen to be in charge of the JS and CSS architecture in my org. I've started trying AI code as a way to write some basic stuff faster (because high-key I have too many responsibilities as it is). I am gobsmacked at HOW FREAKING BLOATED AI CODE IS. And with so many obviously functionality and security holes in a whole swath of things. In other words: I feel ya and stay strong not letting AI add to the tech debt.
0xDEC0DE
Had somebody sloperate a 50-line patch for what was a one-line fix. It introduced a bunch of comments explaining why it was removing lines in tests. So, you know, these would have just ended up being phantom comments explaining literally nothing. Thankfully, it didn't get merged
spartanliam115
im kinda shocked any companies allow AI considering they are leaking all there internal info, code, developments and all the security shit you shouldnt be leaking to a public database.
cousteau
Yeah, the entire tags with negative posts thing is a whole story on its own. Something about some mass-deletion of posts some time in the past having caused it somehow.
spazztastic
IT is stupid hard to get into. I have 2 ex's who's lives and previous jobs were no where close to IT and are now in IT. I've had a background in some of it (most IT jobs don't last long for me) and quickly got a CCNA to get that "for sure" job. Now I'm trying to get a B.S. in compsci but health issues are making it hard
LeftRightThere
Certs get you in the door, at least they used to. I was about to go the networking cert path many years ago because of my military job, but then I discovered I was really good at creating software solutions for infrastructure. There was value in optimizing interoperability in a sea of devs making products that operate in islands.
spazztastic
Ya, I also switched to software. I am about halfway through my engineering degree, but calc 2nd semester is difficult for my brain fog. If I make it to SJSU the CCNA covers a 3rd of the stuff I need to know there lol. I've completed all my coding classes. Used to be military too, USAF Electronic Warfare
InHarmsWay
Tony was really shitty at communicating with Peter in that movie.
LeftRightThere
huey9k
Just like Gen X's parents... hmmm
RooGryphon
ai is for debugging code not writing code
cousteau
Finding random security holes in code is an interesting use of AI I heard about recently, and honestly I don't have any complaints about that, it does seem like a good idea.
somebackup
I work on LLMs. They do better when hyper specialized. Yann is right. Wish he'd hire me, but I'd have to sell first lol. These general purpose LLMs are stupid to use. They still do well with repetitive tasks, like unit tests
bekkayya
depends on what "do better" means. the people paying for and pushing generative models right now are doing it because they have more than zero contextual "understanding" and they can use that to replace the expensive human context component in their business machine. that is what "do better" means to them. a hyper specialized model inherently loses contextual understanding to gain domain specific accuracy so they wouldnt want it, as demonstrated by that we already had machine learning.
somebackup
You're not wrong. The top wants the general purpose. I specialize in domain specific stuff. It just gets you better results less false positives etc. It's still mind numbingly boring in many ways, but it's also more efficient. What I would really like to achieve is something that can learn closer to what a human can. The amount of hours required to teach anything is crazy. A human can learn a lot of it in minutes.
BoobJiggle
Same, I blend and train and test many different llms every week as a hobby. They can be incredibly powerful and useful tools, but context is key. The best way to view an llm is that it's essentially an interactive library index (anybody remember those?).. but the library also has a whole shitload of fiction mixed in with nonfiction so you have to be cautious and validate what it spits out.
I also think "AI" (none of this shit is AI people, ai doesn't exist yet) can be useful finding code bugs
elementcarbon12
Hearing managers say they are making it their mission to push AI and destroy software engineering as a profession makes me sick. But I also kinda want them to push all the engineers away so there is no one but them to be held responsible for when they fail hard.
adiving
Managers that are held responsible? Hah. Golden parachute and next job.
elementcarbon12
These are the low and mid level managers. They have always used workers as an excuse for when things go wrong. The amount of safety the managers have is based on their position in then org chart, their distance from the C suite, and the layers underneath them.
The other bit of irony is how companies have destroyed labor unions but are putting themselves in a position to depend on 1-3 other corporations. They are likely to find those worse negotiations than those with unions.
KiwiGameDev
I’m still writing full manual. When I was younger I wondered what my “old man” thing would be, that aspect of society I outright refused to follow along with no matter how many others said it was fine. This is the hill I’m alright dying on.
skunkworkz
This is something that I struggle with. I'm "old" in this industry (mid '50s). But I haven't stopped learning, haven't stopped being interested in how things work, and haven't stopped improving / modernizing our code base (my main project has been mine since 2003).
But I'm dying on this hill right now. The small amount of good that we might get out of this is so far outweighed by the bad that it causes, that I'm not bringing it into our company at all.
LeftRightThere
In the 90s you could install the internet from a floppy disk. Now, you can download AI models and literally install the knowledge of the internet to your PC.
bekkayya
an impressive feat of compression to be sure but a fact machine it is not. imma opt out of the future where i have to run 5 different AI models to divine if an answer is accurate enough that it might have come from some hand made stack overflow garbage from the beforetimes
KiwiGameDev
I’ve been on the internet long enough to know I don’t want a fraction of what’s around on my computer. Too many feet, wonderbread and snake women eating sonic oc characters
LeftRightThere
Yeah, but when MAGATs start pulling up all the copper in America’s home, you’ll be glad you could search for stuff offline.
Comet260
"juiced autocorrect" As someone that uses AI, that's EXACTLY what it is! GIGO!
Cyanide555
The GPT stands for glorified predictive text
Bear1bear
Ok github is anonymous, share your repos
LeftRightThere
I don’t do the kind of work that gets shared publicly on the internet. I do have some ham radio / hobby electronics projects available.
Bear1bear
Ok, iron man :)
wazeewa
If you want to do it for research, maybe a little bit of troubleshooting, and a soundboard to bounce ideas off of, sure. But this whole vibe coding craze where they just let the text generator do all the work? Not a fucking chance.
0xDEC0DE
I have 20 years of software development experience, working everywhere from literal 5-person startups to Fortune 500 companies. The clankers can write code every bit as good as mine, in a fraction of the time. Given that, why would I even WANT to write code anymore? It's literally a waste of time now. The new skillset that needs to be acquired and cultivated, is keeping the clankers on a VERY short leash in order to maintain quality.
bekkayya
programmer releases scathing takedown of economic system and quality of his peers and his own work, all the while thinking he's bragging lmao
0xDEC0DE
Archaeologist searches among thousands of years of evidence for who the fuck asked you
bekkayya
lol
LeftRightThere
Your experience with AI tools can make better code faster. Where you put your effort is what matters. It’s the transition from developer to engineer, but these new coders aren’t learning this new skills.
Pokegeologist
We have a guy who's self teaching Python and C by having Claude do it all for him.
0xDEC0DE
I taught myself Go that way. I already knew several other programming languages by then, so it was mostly asking the agent questions pointed about unfamiliar syntax and best practices. I initially thought it was super-effective, as I didn't have to waste time going over "this is a variable", "this is a function", and other concepts I was already familiar with, but in hindsight I think it may have stunted my growth
Pokegeologist
I originally started in an engineering program and finished out with a natural science and policy degree. I'm still stuck in the mindset if you can't explain why it works, you don't know.
LeftRightThere
Whenever I get asked what my skills are, I tell people I know how the sausage is made. I know at deep levels how stacks and protocols work. Once you know, you can build anything using that wisdom.
cousteau
I've seen what Claude can do. I'm sorry for that guy. Give him a book.
Recursiv3
I am currently in a battle with my engineering manager to make tickets and work atomic. People keeping opening PRs with 50+ file changes and full features in them. I can't review that. We can't know what that's actually doing. My manager wants to move faster and doesn't see I'm to believe me that the agents are extremely unreliable, ESPECIALLY with complex large contexts 😭
drinkthederpentine
Wow oof. Atomic changes is basic best practice and 2x more important for agents
Ebo352
So, you should document your manager's instance and then respond to them all, cc'ing your skip level, with a link to the company that had its AI delete their database.
Recursiv3
My org VP vibe codes though, I'm pretty much cooked 🚀 but yeah I'm saving emails and screenshots of slack for the inevitable S1/S2 that's coming
TheDreadPirateMontoya
Get your receipts, as the kids say.
thedill2000
Lol ya. Even before AI we'd run into people with huge change sets and a PR two days before release. Then get upset cause the code review is gonna take a week minimum.
Recursiv3
Ahhh yes I remember my junior dev days being like "whew got that in the day before sprint end!" And end of day it has 10+ change requests 🤣
SkittishLittleToaster
What a horror show.
NothingIsSoBad
нехорошо
Ohdearaudrey
Even our architect eventually caved and now we have a “black box” feature because they don’t know how it works, but it accomplishes what they want. It’s just too much to review at once.
Mamaof2QTs
I have accounts typing in requests to chatgpt then passing me the code they want added to my code base. Their prompts are oblivious to the actual data model.
It’s not helpful. I could type crap into chatgpt all day long myself.
0xDEC0DE
It's not helpful at all? In absentia of a formal requirements gathering process, I'd think expressing such things in terms of code might make things easier to reason about. But then again, if they didn't bother to write it, you shouldn't bother to read it. See: https://406.fail/
Kiares
Nightmare fuel
childofsoong
My coworker loves the old "oh, while I'm rewriting the service with a modern http lib, let me also change the formatter config and restructure the build too (and leave a bunch of weird comments to himself) so my diff touches every single file in the repo. If you dare to comment on it, I'll write you messages that are borderline unprofessional"
cousteau
Some people shouldn't be allowed to touch repos. Or actually learn to use them.
sfrinlan
Oh god, fuck that guy. Just right in the dumpster, pick him up and throw him.
childofsoong
Don't get me started - last week I couldn't merge a change to our docs because the build was broken on our dev branch, and I need a passing PR to merge - note that 'docs change' means 'no changes to the code whatsoever'. And when I tried to merge a PR to fix it, I get a rant from him about how I'm stepping on his toes.... he broke the fucking build in the first place when he merged something I had requested updates to (without making the update).
sfrinlan
Did you tell him that breaking the build steps on your toes? Back in the "in office" days, we could just roll our chairs into the person's cube that broke the build and eyeball them until it's fixed. But yeah, the nerve. Breaking the build should be cause for seppuku, not attacking other people.
childofsoong
Yesterday I tried to release something, only to discover he broke our publishing in CI - so I fixed that - and then I found his changes made it publish a skinny jar when I need a fat jar - so I fixed that too...
RocketKokket
I actually got in shit for not writing enough code from a stupid engineering manager a few jobs ago. I was trying to control the tech debt this one junior was just shitting our non stop. She made more features than the rest of the team but she made the entire project unusable by anyone else because of her obtuse over engineered code. I had to spend so much time on each code review just trying to get her to scale things back and got in trouble for not doing as much work as her. SMH
Z0op
Didnt elon fire a bunch of programmers from Xitter just by codelength quotas? And didnt they at one point print out all the code for review?
What idiot is even coming up with the idea for setting that as the standard?
Wishbone1977
I do believe you answered that last question in the first sentence.
Renza0
but I heard blindly trusting AI coding worked great for amazon! /s
SelfconsciouslyConfident
Trader here, using huge protocols with paid models. It cracks me up people think this shit just works. No you have to hold its hand to breathe and then at best its a information tool or a speed tool. Outside of PM work its fine but im a consultant and didnt think much of it. Zoom ai is able to capture everything and follow up if needed. Engineers can handle that
BarryTheCyborg
That's banned at my company as everything we handle is confidential (contract manufacturing) and you don't have any control over what the "AI" companies are doing with the data it's processing.
SelfconsciouslyConfident
I work in Healthcare for a big company you know. There are version that are hippa compliant
bamblebimble
The summaries are pretty useful as I dont write stuff down and then forget what was discussed after a couple days. Granted I rarely use it anyway, typically turning it on when doing an important meeting I can't afford to forget
bekkayya
presumably the summarization leaves out important contextual information at the same rate it makes other errors.
bamblebimble
It might, but seems to do a good enough job to jog my memory when reading it
cousteau
AI can write code, but it can't maintain it. If anything, it can rewrite it. Good luck holding those heaps of vibed code under your responsibility for longer than a month if you're a company that decided this was a good approach :/
LeftRightThere
If your business model is reliability and performance, then you must be deliberate. Making mountains of code isn’t why a project is created. Trusting AI with keys to the kingdom abdicates your responsibility to your customer.
ImJustSoFckingSickOfThisBllsht
It's all about squeezing the AI bubble circle jerk for maximum profits as long as possible until everyone is doing it and enshitification has become the new idiocratic norm.
Comet260
Most MBAs are simply braindead Ketamine snorters.
nihiltres
Enshittification is the hallmark symptom of the organizational illness I’m calling “MBA disease”: dangling your business model over the bleeding edge and seeing how much you can shrink your rope budget. It rarely ends well.
Capitalism can’t tolerate even a stable, profitable status quo; it needs to chase exponential growth *even when that’s impossible,* and then workers get hurt (but *not* leaders) as each bubble of over-leveraged position bursts.
Arbitrarynamehere
They're the absolute dumbest people to leave universities
Hexidimentional
i wouldnt trust anything claude puts out
dirtcreature
Meh. Depends. Prompting is what 99% of devs still think it is. With well formed requirements, keeping memory and sessions religiously, it can do a pretty good job. Skills are also huge. Biggest problem for me is still learning the tool. One thing: never allow auto changes. Ever. Claude will make confident assumptions that lead to disaster, then admit it was an assumption it thought wasn't worth discussing...just like all devs.
Hexidimentional
or hear me out here, i could write the code myself, know its not importing any malicious libraries, and actually know what the methods and variables do.
dirtcreature
Agree with you 100%. But, you can control if you want and it is an amazing tool. Just don't want to see devs lose out because higher up won't listen to reason and think you're in or out if don't ai. Of course, it's fun now...the crying begins in 5 years when no one knows the systems but Claude and it costs too much money to use it.
lowkeybrotherofthor
See I'd. Claude made a badass macro for me for Excel. However, I validated its work before launching it live. Still had to walk it thru what I wanted. Tool is the key word. Drills are tools, but they don't replace the construction worker
Hexidimentional
havent heard of any drills that cause the entire 50,000 residents of lake tahoe to lose their single energy supplier
LeftRightThere
Commercial ones are all about the same if you use them to their strengths. Don’t buy the hype about models being too dangerous to release to the public. So much slop code is being made by AI that other AIs are used to search for vulnerabilities in that slop.
ChiLLeCheeze
I read a news article about Firefox using some AI to find security bugs in FF. Is that all hype? I don't know enough about coding so talk to me like I have a conservative worldview. Any insight into that? https://blog.mozilla.org/en/privacy-security/ai-security-zero-day-vulnerabilities/
LeftRightThere
The part that gets left out is how many iterations are run to find vulnerabilities. Mythos used something like $20,000 in tokens to find a vulnerability that had been found previously using conventional tools.
bekkayya
finding ancient unprotected memcopies is technically a good thing but its not dangerous in the way they're asserting. there have been almost zero exploitable zero days found by AI (almost meaning technically some can be elevated into a problem under specific circumstances where you've already let a bad actor in)
evilspock
The issue isn't writing the code.
The issue is designing the architecture, defining input and output, data and result validation, testing, security, documentation, and deployment/rollback.
The AI can code up a storm, but without the other 85% of the job, it's just an accident waiting for a human to let it happen.
Bad code can be rewritten. But bad design, bad testing, bad security? That's an outage and maybe a fun news story.
MyGreatestFearBoner
Security architect here: this is accurate. Right behind ambitious directors singning contracts with companies that do AI magic, only to be honeydicked by how useless the AI function is.
override367
Jr Developers have taken to producing ten times as much code and Sr engineers can't code at all because they have to review the incredible amount of code. Mid level devs are the most stressed by the changes.
There's an entire skillset to a developer using AI coding tools that isn't being taught anywhere and is being learned in real time: Documentation and Test creation are the most crucial skills
evilspock
The solution there is the AI can't accept responsibility - if a Jr dev makes junk code with the AI, that's on the Jr dev. "It's a poor carpenter who blames his tools" - in the end, the Jr devs who emit crap code stat junior, and those who ]o learn to use the tool correctly succeed.
fiddlewheelx
At worst it would be used for dangerous and critical infrastructure
LeftRightThere
Last year Microsoft had MONTHS of releases that had critical bugs. I suspect that’s when they implemented AI into their pipelines with a “trust me bro” attitude because some manager was impressed with how well copilot replied to emails for them.
evilspock
Maybe? I mean - MS never had a stellar track record. Back in the early 2ks they even devoted a year to "security". Didn't seem to help, but who knows, maybe without it it'd have been even worse. But yes - MS getting sick from drinking their own AI bathwater makes sense.
LeftRightThere
Yeah, but the patches were reliable. We smoke tested updates before pushing them with high confidence. Last year, we were actively withholding some months because of stupid issues or using the previews to mitigate vulnerabilities.
RatsLiveOnNoEvilStar
Thats why I tell people using a LLM to write code is like giving your work to a Jr Dev to develop. Give it clear guidelines and strict paramters and when it give you the output, test it and test it hard before you even think about sending it to the QAs.
bekkayya
"an accident waiting for a human to let happen" yoink
tizkit
don't forget asking questions, i swear the amount of times i finally get a ticket and i ask one question and the product owners have to spend months figuring out the answer is way to high. but if i didn't ask these questions in 3 cases the feature would have been illegal.
evilspock
That's architecture. How can the AI hit a target you can't point it at?
RatsLiveOnNoEvilStar
To me that's just basic software developer stuff. I tell all of my users that 80% of what I do is ask questions. Writing the code to do the stuff is easy. Forcing the users to tell me what the actually want and getting them to think about what their request really means is most of my job.
Loganscomputer
Anyone who successfully uses ai for coding will tell you that you need to recreate the entire pipeline, not just the coder. You need to spin off separate agents for each task so that the context windows stay small. Is the output perfect, not always, but it can be a lot closer.
eetsumkaus
I think though that is exactly where AI can come in, but just like everything, you use it as a tool. high level processes is where vibe coding can help you so you can work out kinks before passing the spec and requirements to the devs. It will help to reduce the cycles of coordination between these sections of the team. You get to work out some of the more obvious bugs before letting another human see it.
evilspock
But that process you describe - and yes, that's exactly what you do - is where the human adds value. The AI is incapable of accurately deciding if its own output is junk - it needs external feedback. With code - it can try running it and get some feedback. If you have good testing, it can get further. But without the scaffolding and feedback from an experienced human - real value is very limited.
Colopty
If you're trying to "work out the bugs" before handing a task over to the devs you're probably doing work outside of scope for your responsibilities and skillset. Just hand the dev work over to the devs while you focus on your actual job, it's far more efficient that way.
eetsumkaus
no, I mean "bugs" in the spec. Things like deadlocks, vulnerabilities, just plain stupid and obvious shit. Once your spec gets big enough then they're inevitable and MOST of the time they're found once you're on the ground and then you have to budget dev time for that. Having a more robust spec to begin with allows you to concentrate on the gnarlier shit that comes up. Or allows you to take on more risk in terms of design.