That is not true, and the proof is that LLMs _can_ reliably generate (relatively small amounts of) working code from relatively terse descriptions. Code is the detail being filled in. Furthermore, LLMs are the ultimate detail fillers, because they are language interpolation/extrapolation machines. And their popularity is precisely because they are usually very good at filling in details: LLMs use their vast knowledge to guess what detail to generate, so the result usually makes sense.
This doesn't detract much from the main point of the article though. Sometimes the interpolated detail is wrong (and indeterministic), so, if reliable result is to be achieved, important details have to be constrained, and for that they have to be specified. And whereas we have decades of tools and culture for coding, we largely don't have that for extremely detailed specs (except maybe at NASA or similar places). We could figure it out in the future, but we haven't yet.
LLMs can generate (relatively small amounts of) working code from relatively terse descriptions, but I don’t think they can do so _reliably_.
They’re more reliable the shorter the code fragment and the more common the code, but they do break down for complex descriptions. For example, try tweaking the description of a widely-known algorithm just a little bit and see how good the generated code follows the spec.
> Sometimes the interpolated detail is wrong (and indeterministic), so, if reliable result is to be achieved
Seems you agree they _cannot_ reliably generate (relatively small amounts of) working code from relatively terse descriptions
In fact, this is a huge chunk of the value a developer brings to the table.
This means that they will very quickly help you discover all the little details that seemed so obvious to you that you didn't even think to mention them, but were nonetheless critical to a successful implementation. The corollary to that is, the potential ROI of outsourcing is inversely proportional to how many of these little details your project has, and how important they are.
So far I've found LLM coding to be kind of the same. For projects where those details are relatively unimportant, they can save me a bunch of effort. But I would not want to let an LLM build and maintain something like an API or database schema. Doing a good job of those requires too much knowledge of expected usage patterns working through design tradeoffs. And they tend to be incredibly expensive to change after deployment so it pays to take your time and get your hands dirty.
I also kind of hate them for writing tests, for similar reasons. I know many people love them for it because writing tests isn't super happy fun times, but for my part I'm tired of dealing with LLM-generated test suites being so brittle that they actively hinder future development.
There is a huge amount of programming work that consists in reinventing the wheel, i.e. in redoing something very similar to programs that have been written thousands of times before.
For this kind of work LLMs can greatly improve productivity, even if they are not much better than if you would be allowed to search, copy and paste from the programs on which the LLM has been trained. The advantage of an LLM is the automation of the search/copy/paste actions, and even more than this, the removal of the copyrights from the original programs. The copyright laws are what has resulted in huge amounts of superfluous programming work, which is necessary even when there are open-source solutions, but the employer of the programmer wants to "own the IP".
On the other hand, for really novel applications, or for old applications where you want to obtain better performance than anyone has gotten before, providing an ambiguous prompt to an LLM will get you nowhere.
The best LLMs can manage is "what's statistically-plausible behaviour for descriptions of humans in the corpus", which is not the same thing at all. Sometimes, I imagine, that might be more useful; but for programming (where, assuming you're not reinventing wheels or scrimping on your research, you're often encountering situations that nobody has encountered before), an alien mind's extrapolation of statistically-plausible human behaviour observations is not useful. (I'm using "alien mind" metaphorically, since LLMs do not appear particularly mind-like to me.)
But I have also caught it on straightforward matters of fact and it’ll apologize. Sometimes in an over the top fashion…
Example: "We should have professional accountability in software"
SWE: "This would bring about the end of the world!!!1!"
this works well for me
It might be illuminating to see what a mathematically precise specification can and cannot do when it comes to generating programs. A major challenge in formal methods is proving that the program implements the specification faithfully, known as the specification gap. If you have a very high level and flexible specification language, such as TLA+, there is a lot of work to do to verify that the program you write meets the specification you wrote. For something like Synquid that is closer to the code there are more constraints on expressivity.
The point is that spoken language is not sufficiently precise to define a program.
Just because an LLM can fill in plausible details where sufficient detail is lacking doesn’t indicate that it’s solving the specification gap. If the program happens to implement the specification faithfully you got lucky. You still don’t actually know that’s true until you verify it.
It’s different with a narrow interface though: you can be very precise and very abstract with a good mathematical system for expressing specifications. It’s a lot more work and requires more training to do than filling in a markdown file and trying to coax the algorithm into outputting what you want through prose and fiction.
Nothing will save you from a bad specification. And there's no royal road to knowing how to write good ones.
you just (correctly) negated your own claim
I'm absolutely on board with that. We probably need less weird and outlier decisions in designs for something that is a boring ass business website.
They can generate boilerplate, sure. Or they can expand out a known/named algorithm implementation, like pulling in a library. But neither of those is generating detail that wasn't there in the original (at most it pulls in the detail from somewhere in the training set).
Only with well-known patterns that represent shared knowledge specified elsewhere. If the details they “fill in” each time differ in ways that change behavior, then the spec is deficient.
If we “figure out” how to write such detailed specs in the future, as you suggest, then that becomes the “code”.
After awhile, I think we all get a sense of not only the amount of micro-decisions you have to make will building stuff (even when you're intimate with the domain), but also the amount of assumptions you'll need to make about things you either don't know yet or haven't fully fleshed out.
I'm painfully aware of the assumptions I'm making nowadays and that definitely changes the way I build things. And while I love these tools, their ability to not only make assumptions, but over-engineer those assumptions can have disastrous effects.
I had Claude build me a zip code heat map given a data source and it did it spectacularly. Same with a route planner. But asking it build out medical procedure documentation configurations based off of a general plan DID NOT work as well as I had expected it would.
Also, I asked Claude about what the cron expression I wrote would do, and it got it wrong (which is expected because Azure Web Jobs uses a non-standard form). But even after telling it that it was wrong, and giving it the documentation to rely on, it still doubled down on the wrong answer.
Not a mere 5 years ago even tech people were chortling down their upturned noses that people were complaining that their jobs were being “taken”, and now that the turns have tabled, there is a bunch of denial, anger, and grief going on, maybe even some depression as many of the recently unemployed realize the current state of things.
It’s all easy to deride the inferiority of AI when you’re employed in a job doing things as you had been all your career, thinking you cannot be replaced… until you find yourself on the other side of the turn that has tabled.
This kind of thing happens at least once per day to me, maybe more.
I am not denying that it is useful, let me be clear. It is extremely convenient, especially for mechanical tasks. It has other advantages like quick exploration of other people's code, for example. If my employer didn't provide a corporate account for me, I would pay one from my own pocket.
That said, I agree with OP and the author that it is not reliable when producing code from specs. It does things right, I would say often. That might be good enough for some fields/people. It's good enough for me, too. I however review every line it produces, because I've seen it miss, often, as well.
It's some nice rhetoric, but you're not actually saying much.
LLMs make developers more efficient. That much is obvious to anyone who isn't blinded by fear.
But people will respond "but you still need developers!" True. You don't need nearly as many, though. In fact, with an LLM in their hands, the poor performers are more of a liability than ever. They'll be let go first.
But even the "smart" developers will be subsumed, as vastly more efficient companies outcompete the ones where they work.
Companies with slop-tolerant architectures will take over every industry. They'll have humans working there. But not many.
This is exactly the argument in Brooks' No Silver Bullet. I still believe that it holds. However, my observation is that many people don't really need that level of details. When one prompts an AI to "write me a to-do list app", what they really mean is that "write me a to-do list app that is better that I have imagined so far", which does not really require detailed spec.
As I mentioned in one of the footnotes in the post:
> People often tell me "you would get better results if you generated code in a more mainstream language rather than Haskell" to which I reply: if the agent has difficulty generating Haskell code then that suggests agents aren't capable of reliably generalizing beyond their training data.
If an agent can't consistently apply concepts learned in one language to generate code in another language, then that calls into question how good they are at reliably permuting the training dataset in the way you just suggested.
Pick a good model, let it choose its own tools and then re-evaluate.
doesn't that apply to flesh-and-bone developers? ask someone who's only working in python to implement their current project in haskell and I'm not so sure you'll get very satisfying results.
No, it does not. If you have a developer that knows C++, Java, Haskell, etc. and you ask that developer to re-implement something from one language to another the result will be good. That is because a developer knows how to generalize from one language (e.g. C++) and then write something concrete in the other (e.g. Haskell).
Code is a description of a solution, which can be executed by a computer. You have the inputs and the outputs (we usually split the former into arguments amd environment, and we split the latter into side effects and return values). Python and Haskell are just different towers of abstraction built on the same computation land. The code may not be the same, but the relation between inputs and outputs does not change if they solve the same problem.
In my experience, a software engineer knows how to program and has experience in multiple languages. Someone with that level of experience tends to pick up new languages very quickly because they can apply the same abstract concepts and algorithms.
If an LLM that has a similar (or broader) data set of languages cannot generalise to an unknown language, then it stands to reason that it is indeed only capable of reproducing what’s already in its training data.
Yes? If they could, we would have a strong general intelligence by now and only few people are claiming this.
I think you’re conflating software and product.
A product can be a recombination of standard software components and yet be something completely new.
This is very true for an email client, but very untrue for an innovative 3D rendering engine technology (just an example).
It necessarily has to derive it from examples of cameras that fly forward that it knows about, without understanding the exact mathematical underpinnings that allow you to rotate a 3D perspective camera and move along its local coordinate system, let alone knowing how to verify whether its implementation functions as desired, often resulting in dysfunctional garbage. Even with a human in the loop that provides it with feedback and grounds it (I tried), it can't figure this out, and that's just a tiny example.
Math is precise, and an LLM's fuzzy approach is therefore a bad fit for it. It will need an obscene amount of examples to reliably "parrot" mathematical constructs.
That's not the task of a renderer though, but its client, so you're talking past your parent comment. And given that I've seen peers one-shot tiny Unity prototypes with agents, I don't really believe they're that bad at taking an educated guess at such a simple prompt, as much as I wish it were true.
I guess the less detailed a spec has to be thanks to the tooling, the more likely it is that the LLM will come up with something usable. But it's unclear to me whether that is because of more examples existing due to higher user adoption, or because of fewer decisions/predictions having to be made by the LLM. Maybe it is a bit of both.
So for me the code is mundane but it’s always unique and rarely do you come across the same problems at different organisations.
If you ever got a spec good enough to be the code, I’m sure Claude or whatever could absolutely ace it, but the spec is never good enough. You never get the context of where your code will run, who will deploy it or what the rollback plan is if it fails.
The code isn’t the problem and never was. The problem is the environment where your code is going.
The proof is bit rot. Your code might have been right 5 years ago but isn’t any more because the world shifted around it.
I am using Claude pretty heavily but there are some problems it is awful at, e.g I had a crusty old classic ASP website to resuscitate this week and it would not start. Claude suggested all the things I half remembered from back in the day but the real reason was Microsoft disabled vbscript in windows 11 24H2 but that wasn’t even on its radar.
I have to remind myself that it’s a fancy xerox machine because it does a damn good job of pretending otherwise.
So goes the AI paradox: it's really effective at writing lots and lots of software that is low value and probably never needed to get written anyway. But at least right now (this is changing rapidly), executives are very willing to hire lots of coders to write software that is low value and probably doesn't need to be written, and VCs are willing to fund lots of startups to automate the writing of lots of software that is low value and probably doesn't need to be written.
Google Maps, Twitter, and AirBnb occupy a tiny fragment of the possible domain applications of software.
Before Google Maps nobody had ever pushed a pure-Javascript AJAX app quite so far; it came out just as AJAX was coined, when user expectations were that any major update to the page required a full page refresh. Indeed, that's exactly what competitor MapQuest did: you had to click the buttons on the compass rose to move the map, it moved one step at a time, and it fully reloaded the page with each move. Google Maps's approach, where you could just drag the map and it loaded the new tiles in the background offscreen, then positioned and cropped everything with Javascript, was revolutionary. Then add that it gained full satellite imagery soon after launch, which people didn't know existed in a consumer app.
Twitter's big innovation was the integration of SMS and a webapp. It was the first microblog, where the idea was that you could post to your publicly-available timeline just by sending an SMS message. This was in the days before Twilio, where there was no easy API for sending these, you had to interface with each carrier directly. It also faced a lot of challenges around the massive fan-out of messages; indeed, the joke was that Twitter was down more than it was up because they were always hitting scaling limits.
HN has (had?) an idiosyncratic architecture where it stores everything in RAM and then checkpoints it out to disk for persistence. No database, no distribution, everything was in one process. It was also written in a custom dialect of Lisp (Arc) that was very macro-heavy. The advantage of this was that it could easily crank out and experiment with new features and new views on the data. The other interesting thing about it was its application of ML to content moderation, and particularly its willingness to kill threads and shadowban users based on purely algorithmic processes.
Claude Code is having the hardest time making sense of it and not breaking everything every step of the way. It always wants to simplify, handwave, "if we just" and "let's just skip if null", it has zero respect for the amount of knowledge and nuance in the product. (Yes, I do have extensive documentation and my prompts are detailed and rarely shorter than 3 paragraphs.)
Most software does something similar. Individual components are pretty simple and well understood, but as you scale your product beyond the simple use cases ("TODO apps"), the interactions between these components create novel challenges. This applies to both functional and non-functional aspects.
So if "cannot make with AI" means "the algorithms involved are so novel that AI literally couldn't write one line of them", then no - there isn't a lot of commercial software like that. But that doesn't mean most software systems aren't novel.
If someone was making a serious request for a to-do list app, they presumably want it to do something different from or better than the dozens of to-do list apps that are already out there. Which would require them to somehow explain what that something was, assuming it's even possible.
I'd pay you 10€ for a TODO app that improved my life meaningfully. It would obviously need to have great UX and be stable. Those are table stakes.
I don't have the time to look at all these apps though. If somebody tells me they made a great TODO app, I'm already mentally filtering them out. There's just too much noise here.
Does your TODO app solve any meaningful problem beyond the bare minimum? Does it solve your procrastination? Does it remind you at the right time?
If it doesn't answer this in the first 2 seconds of your pitch you're out.
There is the problem. Todo apps are easy to make. However making one that is actually useful for tracking todo items it hard. Getting the todo into the app is harder than writing it on paper. Getting reminders at a useful time is hard (now is not a great time to fix that broken widget - it needs parts not in the budget, I'm at work, it needs a couple hours of dedicated time and I have something else coming up...). I've tried a few different ones, most are a combination of too complex and not complex enough at the same time.
Yes; at least, I would hope a musician who was writing a love song was doing so because they want it to do something different from or better than other existing love songs. (Or they might be doing it to practice their songwriting skills - just as a programmer might write a todo app to practice their programming skills - but it makes no sense to use an AI for that)
For some problems, it is. Web front-end development, for example. If you specify what everything has to look like and what it does, that's close to code.
But there are classes of problems where the thing is easy to specify, but hard to do correctly, or fast, or reliably. Much low-level software is like that. Databases, file systems, even operating system kernels. Networking up to the transport layer. Garbage collection. Eventually-consistent systems. Parallel computation getting the same answer as serial computation. Those problems yield, with difficulty, to machine checked formalism.
In those areas, systems where AI components struggle to get code that will pass machine-checked proofs have potential.
But that’s most of the time is not that they want it from objective technical reasons.
They want it because they want to see if they can push you. They do it „because they can”. They do it because later they can renegotiate or just nag and maybe pay less. Multiple reasons that are not technical.
I guess it depends on whether or not we want to make money, or otherwise, compete against others.
The compression ratio is the vibe coding gain.
I think that way of phrasing it makes it easier to think about boundaries of vibe coding.
"A class that represents (A) concept, using the (B) data structure and (C) algorithms for methods (D), in programming language (E)."
That's decodeable, at least to a narrow enough distribution.
"A commercially successful team communication app built around the concept of channels, like in IRC."
Without already knowing Slack, that's not decodable.
Thinking about what is missing is very helpful. Obviously, the business strategic positioning, non technical stakeholder inputs, UX design.
But I think it goes beyond that: In sufficiently complex apps, even purely technical "software engineering" decisions are to some degree learnt from experiment.
This also makes it more clear how to use AI coding effectively:
* Prompt in increments of components that can be encoded in a short prompt.
* If possible, add pre-existing information to the prompt (documentation, prior attempts at implementation).
"Informally, from the point of view of algorithmic information theory, the information content of a string is equivalent to the length of the most-compressed possible self-contained representation of that string. A self-contained representation is essentially a program—in some fixed but otherwise irrelevant universal programming language—that, when run, outputs the original string."
Where it gets tricky is the "self-contained" bit. It's only true with the model weights as a code book, e.g. to allow the LLM to "know about" Slack.
But there is an entire cohort of people who can think about specifying systems but lack the training to sdo so so using the current methods and see a lower barrier to entry in the natural language.
That doesn't mean the LLM is going to think on your behalf (although there is also a little bit of that involved and that's where stuff gets confusing) but it surely provides a completely different interface for turning your ideas into working machinery
"Specifying" is the load-bearing term there. They are describing what they want to some degree, how how specifically?
The uninitiated can continue trying to clumsily refer to the same concepts, but with 100x the tokens, as they lack the same level of precision in their prompting. Anyone wanting to maximize their LLM productivity will start speaking in this unambiguous, highly information-dense dialect that optimizes their token usage and LLM spend...
Setting aside the problem of training, why bother prompting if you’re going to specify things so tightly that it resembles code?
[1] https://en.wikipedia.org/wiki/Lojban
[2] Someone speaking it: https://www.youtube.com/watch?v=lxQjwbUiM9w
Some languages don't have this kind of vocabulary, because there aren't enough speakers that deal with technical things in a given area (and those that do, use another language to communicate)
In my mind this is solving different problems. We want it to parse out our intent from ambiguous semantics because that's how humans actually think and speak. The ones who think they don't are simply unaware of themselves.
If we create this terse and unambiguous language for LLMs, it seems likely to me that they would benefit most from using it with each other, not with humans. Further, they already kind of do this with programming languages which are, more or less, terse and unambiguous expression engines for working with computers. How would we meaningfully improve on this, with enough training data to do so?
I'm asking sincerely and not rhetorically because I'm under no illusion that I understand this or know any better.
Context pollution is a bigger problem.
E.g., those SKILL.md files that are tens of kilobytes long, as if being exhaustively verbose and rambling will somehow make the LLM smarter. (It won't, it will just dilute the context with irrelevant stuff.)
Ah, the Lisp curse. Here we go again.
coincidently, the 80s AI bubble crashed partly because Lisp dialetcts aren't inter-changable.
We literally have proof that an iron age ontology of meaning as represented in Chinese characters is 40% more efficient than naive statistical analysis over a semi phonetic language and we still are acting like more compute will solve all our problems.
Can you elaborate? I think you're talking about https://github.com/PastaPastaPasta/llm-chinese-english , but I read those findings as far more nuanced and ambiguous than what you seem to be claiming here.
Post a link because until you do, I’m almost certain this is pseudoscientific crankery.
Chinese characters are not an “iron age ontology of meaning” nor anything close to that.
Also please cite the specific results in centuries-old “language theory” that you’re referring to. Did Saussure have something to say about LLMs? Or someone even older?
Writing code has always been a means to an end of getting the job done, and I consider being really interested in the practice of coding to be a bit odd. It's as if a carpenter got really into manual saws and got upset about an electric power saw because he "loves manual saws". Whether or not an AI spec is "code" really shouldn't matter if you're a software professional. You are designer/engineer of software systems, not a "coder".
Let’s first agree that code is the final output of a software engineer, and that a wooden object is the final output of a carpenter.
A correct analogy would be, if a software engineer’s AI assisted output is reduced to assembling auto generated code snippets into a workable product, a carpenters is assembling ikea furniture.
The only missing piece is liability. The engineer and carpenter are both hired to assemble something, so they better damn well fully understand what they are delivering because they are being paid to own the liability if anything fails.
LLM -> Spec is easier, especially with good tools that can communicate why the spec fails to validate/compile back to the LLM. Better languages that can codify things like what can actually be called at a certain part of the codebase, or describe highly detailed constraints on the data model, are just going to win out long term because models don't get tired trying to figure this stuff out and put the lego bricks in the right place to make the code work, and developers don't have to worry about UB or nasty bugs sneaking in at the edges.
With a good 'compilable spec' and documentation in/around it, the next LLM run can have an easier time figuring out what is going on.
Trying to create 'validated english' is just injecting a ton of complexity away from the area you are trying to get actual work done: the code that actually runs and does stuff.
Specification means requirements. I like EARS [0] syntax for requirements.
e.g. "while an error is present, the software shall ignore keypresses"
Requirements are not code at all, they are expectations about what the code does; it is the contract about what developers will be held accountable to. Putting implementation details in requirements is a rookie mistake because it takes agency away from the engineers in finding the best solution.
The spec discussed in this article is more akin to the level of detail appropriate in an interface control document (ICD). Very common for a requirement to declare the software shall be compliant to a revision of an ICD.
My own thoughts are: like a good systems engineer that recognizes the software engineers know their domain better than they, we should write specification for AI that leaves room for it to be more clever than we ourselves are. What's the point of exhaustive pseudocoding, it's worse coding. Align on general project preferences, set expectations, and concentrate effort on verifying.
How probable is that given that LLMs output the average of their training data? Your own cleverness would have to be below average for this to hold.
Since every invocation of an LLM may create a different program, just like people, we will see that the spec will leave much room for good and bad implementations, and highlight the imprecision in the spec.
Once we start using a particular implementation it often becomes the spec for subsequent versions, because it's interfaces expose surface texture that other programs and people will begin to rely on.
I'm not sure how well LLMs will fare are brownfield software development. There is no longer a clean specification. Regenerating the code from scratch isn't acceptable. You need TPS reports.
This perfectly explains the feeling I had when, 20 years into my career, I had to start writing specs. I could never quite put my finger on why it was harder than coding. My greater familiarity with coding didn't seem a sufficient explanation.
When writing a line of spec, I had to consider how it might be interpreted in the context of different implementations of other lines of spec - combinatorial nightmare.
But code is just a spec as far as, say, a C compiler is concerned. The compiler is free to implement the assembly however it likes. Writing that spec is definitely easier than writing the assembly (Fred Brookes said this, so it must be true).
So why the difference?
But much of the code we run today is JIT executed, and that leaves ample room for exploiting with weird machines. Eg the TOCTOU in the Corina exploit.
Even at this very low level, full coverage specs require years of careful formal methods work. We have no hope of doing it at for vibe coding, everything will be iterative, and if TDD helps then good, but specs are by no means easier than code.
Not at all. Code is formal, and going from C to assembly os deterministic. While the rules may be complex, they are still rules, and the compiler can’t stray away from them.
Writing C code is easier than writing assembly because it’s an easier notation for thinking. It’s like when hammering a nail. Doing it with a rock is hard. But using a hammer is better. You’re still hitting the nail with a hard surface, but the handle, which is more suitable for a human hand, makes the process easier.
So programming languages are not about the machine performance, they are about easier reasoning. And good programmers can get you to a level above that with proper abstraction and architecture, and give you concepts that directly map to the problem space (Gui frameworks instead of directly rendering to a framebuffer, OS syscall instead of messing with hardware,…).
Natural language is imperfect, code is exact.
The goal of specs is largely to maintain desired functionality over many iterations, something that pure code handles poorly.
I’ve tried inline comments, tests, etc. but what works best is waterfall-style design docs that act as a second source of truth to the running code.
Using this approach, I’ve been able to seamlessly iterate on “fully vibecoded” projects, refactor existing codebases, transform repositories from one language to another, etc.
Obviously ymmv, but it feels like we’re back in the 70s-80s in terms of dev flow.
The latter notion probably is true, but the prior isn’t necessarily true because you can map natural language to strict schemas. ”Implement an interface for TCP in <language>” is probably shorter than the actual implementation in code.
And I understand my example is pedantic, but it extends to any unambiguous definitions. Of course one can argue that TCP spec is not determimistic by nature because natural language isn’t. But that is not very practical. We have to agree to trust some axioms for compilers to work in the first place.
To your point, there are some cases where a short description is sufficient and may have equal or less lines than code (frequently with helper functions utilizing well known packages).
In either case, we’re entering a new era of “compilers” (transpilers?), where they aren’t always correct/performant yet, but the change in tides is clear.
IMHO this could be achieved with large set of tests, but the problem is if you prompt an agent to fix tests, you can't be sure it won't "fix the test". Or implement something just to make the test pass without looking at a larger picture.
When you write software to solve a problem you start with as little as possible details such that when you read it, it would only talk about the business logic. What I mean by that? That you abstract away any other stuff that does not concern the domain you are in, for example your first piece of code should never mention any kind of database technology, it should not contain any reference to a specific communication layer (HTTP for example), and so on.
When you get to this, you have summarized the "spec", and usually it can be read very easily by a non-techincal person (but obviously is a "domain expert") and can also be very testable.
I hope this helps on why the author is right 100%
For a manager, the spec exists in order to create a delgation ticket, something you assign to someone and done. But for a builder, it exists as a thinking tool that evolves with the code to sharpen the understanding/thinking.
I also think, that some builders are being fooled into thinking like managers because ease, but they figure it out pretty quickly.
I guess many of us quality for british parliament.
“With this machine, all the results will be correct, no more errors in log tables”
“And what if people put the wrong figures in?” (Hint - we’d still have the wrong results)
Babbage walks away thinking them an idiot, they walk away thinking Babbage hasn’t considered anything outside of the machine itself.
- Define the data structures in the code yourself. Add comments on what each struct/enum/field does.
- Write the definitions of any classes/traits/functions/interfaces that you will add or change. Either leave the implementations empty or write them yourself if they end up being small or important enough to write by hand (or with AI/IDE autocompletion).
- Write the signatures of the tests with a comment on what it's verifying. Ideally you would write the tests yourself, specially if they are short, but you can leave them empty.
- Then at this point you involve the agent and tell it to plan how to complete the changes without barely having to specify anything in the prompt. Then execute the plan and ask the agent to iterate until all tests and lints are green.
- Go through the agent's changes and perform clean up. Usually it's just nitpicks and changes to conform to my specific style.
If the change is small enough, I find that I can complete this with just copilot in about the same amount of time it would take to write an ambiguous prompt. If the change is bigger, I can either have the agent do it all or do the fun stuff myself and task the agent with finishing the boring stuff.
So I would agree with the title and the gist of the post but for different reasons.
Example of a large change using that strategy: https://github.com/trane-project/trane/commit/d5d95cfd331c30...
I found that to be really vital for good code. https://fsharpforfunandprofit.com/rop/
I’ve found that two to three iterations with various prompts or different models will often yield a surprising solution or some aspect I hadn’t thought of or didn’t know about.
Then I throw away most or all of the code and follow your process, but with care to keep the good ideas from the LLMs, if any.
The only vibecoded thing was an iOS app and I didn't follow this process because I don't know iOS programming nor do I want to learn it. This only works if you know at least how to define functions and data structures in the language, but I think most PMs could learn that if they set their minds to it.
- Do a brainstorming session with the LLM about your idea. Flesh out the major points of the product, who the stakeholders are, what their motivations and goals are, what their pain points are. Research potential competitors. Find out what people are saying about them, especially the complaints.
- Build a high level design document with the LLM. Go through user workflows and scenarios to discern what kinds of data are needed, and at what scale.
- Do more market research to see how novel this approach is. Figure out what other approaches are used, and how successful they are. Get user pain points with each approach if you can. Then revisit your high level design.
- Start a technical design document with the LLM. Figure out who the actors of the system are, what their roles are in the system, and what kinds of data they'll need in order to do their job.
- Research the technologies that could help you build the system. Find out how popular they are, how reliable they are, how friendly they are (documentation, error messaging, support, etc), their long-term track record, etc. These go into a research document.
- Decide based on the research which technologies match your system best. Start a technical document with the LLM. Go through the user scenarios and see how the technologies fit.
- Decide on the data structures and flows through the system. Caching, load balancing, reliability, throughput requirements at the scale you plan to reach for your MVP and slightly beyond. Some UX requirements at this point are good as well.
- Start to flesh out your interfaces, both user and machine. Prototype some ideas and see how well they work.
- Circle back to research and design based on your findings. Iterate a few times and update the documents as you go using your LLM. Try to find ways to break it.
- Once you're happy with your design, build an architecture document that shows how the whole system will concretely fit together.
- Build an implementation plan. Run it through multiple critique rounds. Try to find ways to break it.
- Now you're at the threshold where changes get harder. Build the implementation piece by piece, checking to make sure they work as expected. This can be done quickly with multiple LLMs in parallel. Expect that the pieces won't fit and you'll need to rethink a lot of your assumptions. Code will change a LOT, so don't waste much time making it nice. You should have unit and integration tests and possibly e2e tests, which are cheap for the LLM to maintain, even if a lot of them suck.
- Depending on how the initial implementation went, decide whether to keep the codebase and refine it, or start the implementation over using the old codebase for lessons learned.
Basically, more of the same of what we've been doing for decades, just with faster tools.
I have always done that steps with my team and the results are great.
If you are a solo developer I understand that the LLM can help somewhat but not replace a real team of developers.
Those very detailed specs then let agents run for a long time without supervision so nice for multi tasking :)
The code will always be an imperfect projection of the specification, and that is a feature. It must be decoupled to some extent or everything would become incredibly brittle. You do not need your business analysts worrying about which SQLite provider is to be used in the final shipped product. Forcing code to be isomorphic with spec means everyone needs to know everything all the time. It can work in small tech startups, but it doesn't work anywhere else.
Most people don't specify or know that they want a U bend in their pipes or what kind or joints should be used for their pipes.
The absence of U bends or use or poor joints will be felt immediately.
Thankfully home building is a relatively solved problem whereas software is everything bespoke and every problem slightly different... Not to mention forever changing
> For every specification satisfied by the input program, the output program satisfies the same specification.
This is not a program and it does not become a program once you fill in the holes. Making the statement precise clearly requires a formal language, but that language can work at a higher level of abstraction than a programming language. So yes, a specification can absolutely be simpler than a program that implements it.
Anyone who studied software engineering, should know that specification doesn’t bother with implementation details of the underlying technology.
Things such as quite specific engine are used, are the contents of an encapsulated subsystem.
Proper software engineering specification is incompatible with a hacker culture and picking technology beforehand is a bad practice. It’s much closer to waterfall than to C4.
However, the last 20 years we got software building blocks which impose system architectural restrictions: frameworks. And also pieces of software which are half cooked systems.
Far are the days of requirements, preconditions, postconditions and invariants, network diagrams and entity relationship models.
We have spent decades working on reproducible builds or deterministic compilation. To achieve this, all steps must be deterministic. LLMs are not deterministic. You need to commit source code.
I used to be on that side of the argument - clearly code is more precise so it MUST be simpler than wrangling with the uncertainty of prose. But precision isn't the only factor in play.
The argument here is that essential complexity lives on and you can only convert between expressions of it - that is certainly true but it's is overlooking both accidental complexity and germane complexity.
Specs in prose give you an opportunity to simplify by right-sizing germane complexity in a way that code can't.
You might say "well i could create a library or a framework and teach everyone how to use it" and so when we're implementing the code to address the essential complexity, we benefit from the germane complexity of the library. True, but now consider the infinite abstraction possible in prose. Which has more power to simplify by replacing essential complexity with germane complexity?
Build me a minecraft clone - there's almost zero precision here, if it weren't for the fact that word minecraft is incredibly load bearing in this sentence, then you'd have no chance of building the right thing. One sentence. Contrast with the code you'd have to write and read to express the same.
Technical design docs are higher level than code, they are impricise but highlight an architectural direction. Blanks need to be filled in. AI Shines here.
Formal specs == code Some language shine in being very close to a formal spec. Yes functional languages.
But lets first discuss which kind of spec we talk about.
- Delete code and start all over with the spec. I don't think anyone's ready to do that.
- Buy a software product / business and be content with just getting markdown files in a folder.
I haven't heard of being content paying for a product consisting of markdown files. Though I could imagine people paying for agent skill files. But yet, the skills are not the same product as say, linear.
This article is really attacking vague prose that pushes ambiguity onto the agent - okay, fair enough. But that's a tooling problem. What if you could express structure and relationships at a higher level than text, or map domain concepts directly to library components? People are already working on new workflows and tools to do just that!
Also, dismissing the idea that "some day we'll be able to just write the specs and the program will write itself" is especially perplexing. We're already doing it, aren't we? Yes, it has major issues but you can't deny that AI agents are enabling literally that. Those issues will get fixed.
The historical parallel matters here as well. Grady Booch (co-creator of UML) argues we're in the third golden age of software engineering:
- 1940s: abstracted away the machine -> structured programming
- 1970s: abstracted away the algorithm -> OOP, standard libraries, UML
- Now: abstracting away the code itself
Recent interview here: https://www.youtube.com/watch?v=OfMAtaocvJw
Each previous transition had engineers raising the same objections: "this isn't safe", "you're abstracting away my craft". They were right that something was lost, but wrong that it was fatal. Eventually the new tools worked well enough to be used in production.
Which was mostly a failure, to the point that there is a major movement towards languages that "abstract away" (in this sense) less, e.g. Rust.
Certainly if the creators of UML are saying that AI is great, that gives me more confidence than ever that it's bunk.
Rust uses crates to import those dependencies, which was one of its biggest innovations.
A sufficiently detailed spec need only concern itself with essential complexity.
Applications are chock-full of accidental complexity.
That is the great insight I can offer
Looks like I would have been better off!
Then came all sorts of shenanigans, from memory management to syntax hell, which took forever to learn effectively.
This stage was a major barrier to entry, and it's now gone — so yeah, things have indeed changed completely.
I sort of see the success of coding agents as not just a sign that people get tricked by slop, but mainly a sign that today's languages and frameworks just aren't that good. And now it's easier to say to the LLM, "I literally don't care how the UI is organized as long as everything is visible" than to try to express that in code.
[1]: I mean this in the vaguest sense, so including DSLs/frameworks/interfaces offered by libraries too
Like they say “everything comes round again”
“This time is different” are the famous last words.
Could be that the truth is somewhere in between?
Simple thought experiment: Imagine you have a spec and some code that implements it. You check it, and find that a requirement in the spec was missed. Obviously that code is not the spec; the spec is the spec. But also you couldn't even use the code as a spec because it was wrong. Now remove the spec.
Is the code a spec for what was supposed to be built? No. A requirement was missed. Can you tell from just the code? Also no. You need a two separate sources that tell you what was meant to be written in case the either of them is wrong. That is usually a spec and the code.
They could both be wrong, and often are, but that's a people problem.
Or the spec was wrong and the code was right; that happens more often than you might think. If we view it through the lens of "you have two specs, one of which is executable" then obviously you could have made an error in either spec, and obviously writing two specs and checking for mismatches between them is one possible way to increase the fidelity of what you wrote, but it's nothing more than that.
I feel like if you’re designing a language, the activity of producing the spec, which involves the grammar etc., would allow you to design unencumbered by whether your design is easy to implement. Or whether it’s a good fit for the language you are implementing the compiler with.
The OP also correctly identifies that thoughtful design takes a back seat in favor of action when we start writing the code.
So unless you want bugs to be your specification, you actually need to specify what you want.
If you want a specification from source code, you need to reverse engineer it. Although that’s a bit easier now, with LLMs.
Pure specification itself is useless without actual implementation (which does the job), so, trying to write such specification (in a natural language) has no practical sense.
A specification describes the externally observable behavior, constraints, and properties of a system. That's independent of how those properties are implemented.
An implementation is a concrete realization of a specification, that achieves the specified behavior using particular algorithms, data structures, and operational mechanisms.
> Pure specification itself is useless without actual implementation (which does the job), so, trying to write such specification (in a natural language) has no practical sense.
That's a non sequitur, i.e. the part of your sentence after the commas doesn't follow from the first part.
"Pure specification itself" has a purpose, which is to guide implementations. That's the "practical sense".
If you look at what implementions modern compilers actually come up with, they often look quite different from what you would expect from the source code
If your program was a DSL for steering, the abstract machine will be the idea of steering wheel, while the machine could be a car without one. So a compiler would build the steering wheel, optionally adding power steering (optimization), and then tack the apparatus to steer for the given route.
I also enjoyed the writing style so much that I felt bad for myself for not getting to read this kind of writing enough. We are drowning in slop. We all deserve better!
fn sin(x: f16) -> f16
Yet sin() here can have a large number of different implementations. The spec alone under-determines the actual code.
If "spec" doesn't imply that, what does it mean to you? Or maybe you are suggesting that C "code" isn't code[1] either?
[1] By the original definition of code that is actually quite true, but I think we can agree the term, as normally used, has evolved over the years?
To invert your polynomial analogy, the specification might call for a sine wave, your code will generate a Taylor series approximation that is computable.
Simply put: Formal language = No ambiguities.
Once you remove all ambiguous information from an informal spec, that, whatever remains, automatically becomes a formal description.
Or everything will converge toward a rust like inference optimized gibberish...
"Is this it?" "NOPE"
That is simply not true. There is a ton of litterature around inherent vs accidental complexity, which in an ideal world should map directly to spec vs code. There are a lot of technicalities in writing code that a spec writer shouldn't know about.
Code has to deal with the fact that data is laid out a certain way in ram and on disk, and accessing it efficiently requires careful implementation.
Code has to deal with exceptions that arise when the messiness of the real world collides with the ideality of code.
It half surprises me that this article comes from a haskell developer. Haskell developers (and more generally people coming from maths) have this ideal view of code that you just need to describe relationships properly, and things will flow from there.
This works fine up to a certain scale, where efficiency becomes a problem.
And yes, it's highly probable that AI is going to be able to deal with all the accidental complexity. That's how I use it anyways.
I did a side project with a non-technical co-founder a year ago and every time he told me what he wanted, I made a list of like 9 or 10 logical contradictions in his requirements and I had to walk him through what he said with drawings of the UI so that he would understand. Some stuff he wanted me to do sounded good in his head but once you walk through the implementation details, the solution is extremely confusing for the user or it's downright physically impossible to do based on cost or computational resource constraints.
Sure, most people who launched a successful product basically stumbled onto the perfect idea by chance on the first attempt... But what about the 99% others who fell flat on their face! You are the 99% and so if you want to succeed by actual merit, instead of becoming a statistic, you have to think about all this stuff ahead of time. You have to simulate the product and business in detail in your mind and ask yourself honestly; is this realistic? Before you even draw your first wireframe. If you find anything wrong with it, anything wrong at all; it means the idea sucks.
It's like; this feature is too computationally and/or financially expensive to offer for free and not useful enough to warrant demanding payment from users... You shouldn't even waste your time with implementation; it's not going to work! The fundamental economics of the software which exists in your imagination aren't going to magically resolve themselves after implementing in reality.
Translating an idea to reality never resolves any known problems; it only adds more problems!
The fact is that most non-technical people only have a very vague idea of what they want. They operate in a kind of wishy washy, hand-wavy emotion-centric environment and they think they know what they're doing but they often don't.
When I code by hand under time pressure, I’m actually more likely to dig a hole. Not because I can’t write code, but because humans get impatient, bored, optimistic and sloppy in predictable ways. The machine doesn’t mind doing the boring glue work properly.
But that is not the real problem.
The real problem is what happens when an organisation starts shipping code it does not understand. That problem predates LLMs and it will outlive them. We already live in a world full of organisations that ship bad systems nobody fully understands, and the result is the usual quagmire: haunted codebases, slow change, fear-driven development, accidental complexity, and no one knowing where the actual load-bearing assumptions are.
LLMs can absolutely make that worse, because they increase the throughput of plausible code. If your bottleneck used to be code production, and now it’s understanding, then an organisation that fails to adapt will just arrive at the same swamp faster.
So to me the important distinction is not “spec vs code”. It’s more like:
• good local code is not the same thing as system understanding
• passing tests are not the same thing as meaningful verification
• shipping faster is not the same thing as preserving legibility
The actual job of a programmer was never just to turn intent into syntax anyway. Every few decades the field reinvents some story about how we no longer need programmers now: Flow-Matic, CASE tools, OO, RUP, low-code, whatever. It’s always the same category error. The hard part moves up a layer and people briefly mistake that for disappearance.
In practice, the job is much closer to iteratively solving a problem that is hard to articulate. You build something, reality pushes back, you discover the problem statement was incomplete, the constraints were wrong, the edge case was actually central, the abstraction leaks, the user meant something else, the environment has opinions, and now you are solving a different problem than the one you started with.
That is why I think the truly important question is not whether AI can write code.
It’s whether the organisation using it can preserve understanding while code generation stops being the bottleneck.
If not, you just get the same bad future as before, only faster, cleaner-looking, and with more false confidence.
The blue print analogy comes up frequently. IMHO this is unfortunate. Because a blueprint is an executable specification for building something (manually typically). It's code in other words. But for laborers, construction workers, engineers, etc. For software we give our executable specifications to an interpreter or compiler. The building process is fully automated.
The value of having specifications for your specifications is very limited in both worlds. A bridge architect might do some sketches, 3D visualizations, clay models, or whatever. And a software developer might doodle a bit on a whiteboard, sketch some stuff out on paper or create a "whooooo we have boxes and arrows" type stuff in a power point deck or whatever. If it fits on a slide, it has no meaningful level of detail.
As for AI. I don't tend to specify a lot when I'm using AI for coding. A lot of specification is implicit with agentic coding. It comes from guard rails, implicit general knowledge that models are trained one, vague references like "I want red/green TDD", etc. You can drag in a lot of this implicit stuff with some very rudimentary prompting. It doesn't need to be spelled out.
I put an analytics server live a few days ago. I specified I wanted one. And how I wanted it to work. I suggested Go might be a nice language to build it in (I'm not a Go programmer). And I went in to some level of detail on where and how/where I wanted the events to be stored. And I wanted a light js library "just like google analytics" to go with it. My prompt wasn't much larger than this paragraph. I got what I asked for and with some gentle nudging got it in a deployable state after a few iterations.
A few years ago you'd have been right to scald me for wasting time on this (use something off the shelf). But it took about 20 minutes to one shot this and another couple of hours to get it just right. It's running live now. As far as I can see with my few decades of experience, it's a pretty decent version of what I asked for. I did not audit the code. I did ask for it to be audited (big difference) and addressed some of the suggested fixes via more prompting ("sounds good, do it").
If you are wondering why, I'm planning to build a AI dashboard on top of this and I need the raw event store for that. The analytics server is just a dirt cheap means to an end to get the data where I need it. AI made the server and the client, embedded the client in my AI generated website that I deployed using AI. None of this involved a lot of coding or specifying. End to end, all of this work was completed in under a week. Most of the prompting work went into making the website really nice.
Plan mode is a trap. It makes you feel like you're actually engineering a solution. Like you're making measured choices about implementation details. You're not, your just vibe coding with extra steps. I come from an electrical engineering background originally, and I've worked in aerospace most of my career. Most software devs don't know what planning is. The mechanical, electrical, and aerospace engineering teams plan for literal years. Countless reviews and re-reviews, trade studies, down selects, requirement derivations, MBSE diagrams, and God knows what else before anything that will end up in the final product is built. It's meticulous, detailed, time consuming work, and bloody expensive.
That's the world software engineering has been trying to leave behind for at least two decades, and now with LLMs people think they can move back to it with a weekend of "planning", answering a handful of questions, and a task list.
Even if LLMs could actually execute on a spec to the degree people claim (they can't), it would take as long to properly define as it would to just write it with AI assistance in the first place.
LLMs are very good at bash, which I’d argue doesn’t read like language or math.
Posting something like this in 2026: they must not have heard of LLMs.
And also: this is such a typical thing a functional programmer would say: for them code is indeed a specification, with strictly no clue (or a vague high level idea at most) as to how the all effing machine it runs on will actually conduct the work.
This is not what code is for real folks with real problems to solve outside of academic circles: code is explicit instructions on how to perform a task, not a bunch of constraints thrown together and damned be how the system will sort it out.
And to this day, they still wonder why functional programming almost never gets picked up in real world application.
Secondly, they are able to produce intelligence that wasn't represented in their training input. As a simple example take a look at chess AI. The top chess engines have more intelligence over the game of chess than the top humans. They have surpassed humans understanding of chess. Similar with LLMs. They train on synthetic data that other LLMs have made and are able to find ways to get better and better on their own. Humans learn off the knowledge of other humans and it compounds. The same thing applies to AI. It is able to generated information and try things and then later reference what it tried when doing something else.
That was partly possible because chess is a constrained domain: rigid rules and board states.
But LLM land is not like that. LLM land was trained on pre-existing text written by humans. They do discover patterns within said data but the point stands that the data and patterns within are not actually novel.
Some of the pretraining. Other pretraining is on text written by AI. Human training data is only but a subset of what these models train on. There is a ton of synthetic training data now.
The difficulty has always arisen when the lines of code pile up AND users start requesting other things AND it is important not to break the "unintended behavior" parts of the system that arose from those initial decisions.
It would take either a sea-change in how agents work (think absorbing the whole codebase in the context window and understanding it at the level required to anticipate any surprising edge case consequences of a change, instead of doing think-search-read-think-search-read loops) or several more orders of magnitude of speed (to exhaustively chase down the huge number of combinations of logic paths+state that systems end up playing with) to get around that problem.
So yeah, hobby projects are a million times easier, as is bootstrapping larger projects. But for business works, deterministic behaviors and consistent specs are important.
They aren't significantly better now than a couple of years ago. So it doesn't seem likely they will be significantly better in a couple of years than they are now.
Say you and I both wrote the same spec that under-specifies the same parts. But we both expect different behavior, and trust that LLM will make the _reasonable_ choices. Hint: “The choice that I would have made.”
Btw, by definition, when we under-specify we leave some decisions to the LLM unknowingly.
And absent our looks or age as input, the LLM will make some _reasonable_ choices based on our spec. But will those choices be closer to yours or mine? Assuming it won’t be neither.