Cloudlflare builds OAuth with Claude and publishes all the prompts

If you look at the README it is completely revealed... so i would argue there is nothing to "reveal" in the first place.

> I started this project on a lark, fully expecting the AI to produce terrible code for me to laugh at. And then, uh... the code actually looked pretty good. Not perfect, but I just told the AI to fix things, and it did. I was shocked.

> To emphasize, this is not "vibe coded". Every line was thoroughly reviewed and cross-referenced with relevant RFCs, by security experts with previous experience with those RFCs.

I hate to say, though, but I have reviewed a lot of human code in my time, and I've definitely caught many humans making similar-magnitude mistakes. :/

removing expert humans from the loop is the deeply stupid thing the Tech Elite Who Want To Crush Their Own Workforces / former-NFT fanboys keep pushing, just letting an LLM generate code for a human to review then send out for more review is really pretty boring and already very effective for simple to medium-hard things.

I estimate it would have taken a few weeks, maybe months to write by hand.

That said, this is a pretty ideal use case: implementing a well-known standard on a well-known platform with a clear API spec.

In my attempts to make changes to the Workers Runtime itself using AI, I've generally not felt like it saved much time. Though, people who don't know the codebase as well as I do have reported it helped them a lot.

I have found AI incredibly useful when I jump into other people's complex codebases, that I'm not familiar with. I now feel like I'm comfortable doing that, since AI can help me find my way around very quickly, whereas previously I generally shied away from jumping in and would instead try to get someone on the team to make whatever change I needed.

The million dollar question is, what are the unintended, unpredicted consequences of developing this way?

If AI allows me to write code 10x faster, I might end up with 10x more code. Has our ability to review it gotten equally fast? Will the number of bugs multiply? Will there be new classes of bugs? Will we now hire 1 person where we hired 5 before? If that happens, will the 1 person leaving the company become a disaster? How will hiring work (cuz we have such a stellar track record at that...)? Will the changing economics of creating software now make SaaS no longer viable? Or will it make traditional commercial software companies no longer viable? Will the entire global economy change, the way it did with the rise of the first tech industry? Are we seeing a rebirth?

We won't know for sure what the consequences are for a while. But there will be consequences.

But what if you only need 2 kentonv's instead of 20 at the end? Do you assume we'll find enough new tasks that will occupy the other 18? I think that's the question.

And the author is implementing a fairly technical project in this case. How about routine LoB app development?

I *think* the answer to this is clearly no: or at least, given what we can accomplish today with the tools we have now, and that we are still collectively learning how to effectively use this, there's no way it won't be faster (with effective use) in another 3-6 months to fully-code new solutions with AI. I think it requires a lot of work: well-documented, well-structured codebases with fast built-in feedback loops (good linting/unit tests etc.), but we're heading there no

My problem is that (in my experience anyways) this is slower than me just writing the code myself. That's why AI is not a useful tool right now. They only get it right sometimes so it winds up being easier to just do it yourself in the first place. As the saying goes: bad help is worse than no help at all, and AI is bad help right now.

There is a middle ground: software engineers being kicked out because now some business person can hand over the task of building the entire OAuth infrastructure to a single inexperienced developer with a Claude account.

And where are supposed to get experienced engineers if replaced all Jr Devs with AI? There is a ton of benefit from drudgery of writing classes even if seems like grunt work at the time.

The theory of enshittification says that "business person pressing a few buttons" approach will be pursued, even if it lowers quality, to save costs, at least until that approach undermines quality so much that it undermines the business model. However, nobody knows how much quality tradeoff tolerance is there to mine.

Why would a human review the code in a few years when AI is far better than the average senior developer? Wouldn't that be as stupid as a human reviewing stockfish's moves in Chess?

If a robot assembles cars at lightning speed... but occasionally misaligns a bolt, and your only safeguard is a visual inspection afterward, some defects will roll off the assembly line. Human coders prevent many bugs by thinking during assembly.

Side note, reverse engineering SAP ABAP sounds torturous.

First you use any LLM with a large context to write down the plan - preferably in a markdown file with checkboxes "- [ ] Task 1"

Then you can iterate on the plan and ask another LLM more focused on the subject matter to do the tasks one by one, which allows it to work without too much hallucination as the context is more focused.

https://github.com/cloudflare/workers-oauth-provider/blob/fe...

LLM assisted coding is a way to get stuff done much faster, at a greatly increased mental cost / energy spent. Oddly enough.

How are you using it?

I've been mainly doing "pair programming" with my own agent (using Devstral as of late) and find the reviewing much easier than it would been to literally type all of the code it produces, at least time wise.

I've also tried vibe coding for a bit, and for that I'd agree with you, as you don't have any context if you end up wanting to review something. Basically, if the project was vibe coded from the beginning, it's much harder to get into the codebase.

But when pair programming with the LLM, I already have a built up context, and understand how I want things to be and so on, so reviewing pair programmed code goes a lot faster than reviewing vibe coded code.

Funnily, enough, I find the exact opposite. I feel so much relief that I don't have to waste time figuring out every, single detail. It frees me up to focus on architectural and higher level changes.

Restart the conversation from scratch. As soon as you get something incorrect, begin from the beginning.

It seems to me like any mistake in a messages chain/conversation instantly poisons the output afterwards, even if you try to "correct" it.

So if something was wrong at one point, you need to go back to the initial message, and adjust it to clarify the prompt enough so it doesn't make that same mistake again, and regenerate the conversation from there on.

My point being: assuming you have RFCs (which leave A LOT to the imagination) and some OSS implementations to train on, each implementation usually has too many highly specific choices made to safely assume an LLM would be able to cobble something together without an amount of oversight effort approaching simply writing the damned thing yourself.

A few months ago, solving such a spec riddle could take a while, and most of the time, the solutions that were produced by long run times were worse than the quick solutions. However, recently the models have become significantly better at solving such riddles, making it fun (depending on how well your use case can be put into specs).

In my experience, sonnet 3.7 represented a significant step forward compared to sonnet 3.5 in this discipline, and Gemini 2.5 Pro was even more impressive. Sonnet 4 makes even fewer mistakes, but it is still necessary to guide the AI through sound software engineering practices (obtaining requirements, discovering technical solutions, designing architecture, writing user stories and specifications, and writing code) to achieve good results.

Edit: And there is another trick: Provide good examples to the AI. Recently, I wanted to create an app with the OpenAI Realtime API and at first it failed miserably, but then I added the most important two pages of the documentation and one of the demo projects into my workspace and just like that it worked (even though für my use-case the API calls had to be use quite differently).

Eventually you'll build up a somewhat reusable template you can use as a system prompt to guide it exactly how you want.

Basically, you get what you ask for, nothing else and nothing more. If you're unclear, it'll produce unclear outputs, if you didn't mention something, it'll do whatever with that. You have to be really, really explicit about everything.

But there have been many cases in my experience where the LLM could not possibly have been simply pattern-matching to something it had seen before. It really did "understand" the meaning of the code by any definition that makes sense to me.

Relevant post: https://news.ycombinator.com/item?id=44089156

Then you should be able to make a markov chain generator without deep neural nets, and it should be on the same level of performance as current LLMs.

But we both know you can't.

As soon as compression happens, optimization happens which can lead to rules/learning of principles which got feed by statistics.

This isn't vibecoding. This is LLM-assisted coding.

I absolutely would not vibe code an OAuth implementation! Or any other production code at Cloudflare. We've been using more AI internally, but made this rule very clear: the human engineer directing the AI must fully understand and take responsibility for any code which the AI has written.

I do think vibe coding can be really useful in low-stakes environments, though. I vibe-coded an Android app to use as a baby monitor (it just streams audio from a Unifi camera in the kid's room). I had no previous Android experience, and it would have taken me weeks to learn without AI, but it only took a few hours with AI.

I think we are in desperate need of safe vibe coding environments where code runs in a sandbox with security policies that make it impossible to screw up. That would enable a whole lot of people to vibe-code personal apps for personal use cases. It happens I have some background building such platforms...

But those guardrails only really make sense at the application level. At the systems level, I don't think this is possible. AI is not smart enough yet to build systems without serious bugs and security issues. So human experts are still going to be necessary for a while there.

I don't know if it was the intent but these kind of questions bother me, the seem to hint at an agenda, "when can I have a farm of idiots with keyboards paid minimum wage churn out products indistinguishable from expertly designed applications".

To me that's the danger of AI, not it's purported intelligence, but our manifested greed.

i.e. I might not use AI to build an OAuth library, but I might use AI to build a web app (which I am an expert at) that may use an OAuth library Cloudfare developed (which theya are experts at). Trying to make "anyone" code "anything" doesn't seem like the point to me

My 2 cents:

>I guess for me the questions is, at what point do you feel it would be reasonable to this without the experts involved in your case?

No sooner and no later than we could say the same thing about a junior developer. In essence, if you can't validate the code produced by a LLM then you shouldn't really have been writing that code to begin with.

>The really really interesting thing would be if an AI could actually generate the prompts.

I think you've hit on something that is going underexplored right now in my opinion. Orchestration of AI agents, where a we have a high level planning agent delegating subtasks to more specialized agents to perform them and report back. I think an approach like that could help avoid context saturation for longer tasks. Cline / Aider / Roo Code / etc do something like this with architect mode vs coding mode but I think it can be generalized.

Depends on the project. Word on the street is the closer your project is to an archetypical React tutorial TODO App then you'll likely be pleased with the results. Whereas if your project is a WDK driver in Rust where every file is a minefield then you'll spend the next few evenings having to audit everything with a fine toothed comb.

> since AI will rewrite it

That depends if you believe in documentation-first or program-first definitions of a specification.

Start at the bottom...they are in the commit messages, or sometimes the .md file

This is intended for building lightweight services quickly. Historically there has been no real need for "lightweight" OAuth providers -- if you were big enough that people wanted to connect to you using OAuth, you were not lightweight. MCP has sort of changed that as the "big" side of an MCP interaction is the client side (the LLM provider), whereas lots of people want to create all kinds of little MCP servers to do all kinds of little things. But MCP specifies OAuth as the authentication mechanism. So now people need to be able to implement OAuth from the provider side easily.

> Experience with the RFCs but have not been able to correctly name it.

These docs are written for people building MCP servers, most of whom only know they want to expose an API to AIs and have never read OAuth RFCs. They do not know or care about the difference between an authorization server and a resource server.

Strictly speaking, yes. But speaking of IDPs more broadly, it’s perfectly acceptable to refer to the authorisation-server as an auth-provider, especially in OIDC (which is OAuth, with extensions) where it’s explicitly called “OpenID provider” - so it’s natural for anyone well-versed in both to cross terminology like that.

That said, if you have spotted a place in the code where you believe there is such a vulnerability, please do report it. Disclosure guidelines are at: https://github.com/cloudflare/workers-oauth-provider/blob/ma...

For example, I usually come off as being relatively skeptic within the HN crowd, but I'm actually pushing for more usage at work. This kind of "opinion arbitrage" is common with new technologies.

This is my problem I run into quite frequently. I have more trouble trying to explain computing or architectural concepts in natural language to the AI than I do just coding the damn thing in the first place. There are many reasons we don't program in natural language, and this is one of them.

I've never found natural language tools easier to use, in any iteration of them, and so I get no joy out of prompting AI. Outside of the increasingly excellent autocomplete, I find it actually slows me down to try and prompt "correctly."

Most people who are good at a skill start here with AI. Over time, they learn how to explain things better to AI. And the output from AI improves significantly as a result. Additionally AI models keep improving over time as well.

If you stay with it, you will reach an inflection point soon enough and will actually start enjoying it.

Of course, we won't be able to tell the real effects, now, because every longitudinal study of researchers will now be corrupted by the ongoing evisceration of academic research in the current environment. Vibe-coding won't be a net creativity gain to a researcher affected by vibe-immigration-policy, vibe-grant-availability, and vibe-firings, for all of which the unpredictability is a punitive design goal.

Whether fear of LLMs taking jobs has contributed to a larger culture of fear and tribalism that has emboldened anti-intellectual movements worldwide, and what the attributable net effect on research and development will be... it's incredibly hard to quantify.

* We achieved Level 2 autonomy first, which requires you to fully supervise and retain control of the vehicle and expect mistakes at any moment. So kind of neat but also can get you in big trouble if you don't supervise properly. Some people like it, some people don't see it as a net gain given the oversight required.

^ This is where Tesla "FSD beta" is at, and probably where LLM codegen tools are at today.

* After many years we have achieved a degree of Level 4 autonomy on well-trained routes albeit with occasional human intervention. This is where Waymo is at in certain cities. Level 4 means autonomy within specific but broad circumstances like a given area and weather conditions. While it is still somewhat early days it looks like we can generally trust these to operate safely and ask for help when they are not confident. Humans are not out of the loop.[1]

^ This is probably what where we can expect codegen to grow after many more years of training and refinement in specific domains. I.e. a lot of what CloudFlare engineers did with their prompt engineering tweaking was of this nature. Think of them as the employees driving the training vehicles around San Francisco for the past decade. And similarly, "L4 codegen" needs to prioritize code safety which in part means ensuring humans can understand situations and step in to guide and debug when the tool gets stuck.

* We are still nowhere close to Level 5 "drive anywhere and under any conditions a human can." And IMHO it's not clear we ever will based purely on the technology and methods that got us to L4. There are other brain mechanisms at work that need to be modeled.

[1] https://www.cnbc.com/2023/11/06/cruise-confirms-robotaxis-re...

Does it even have to be able to do so? Just the ability to speed up exploration and validation based on what a human tells it to do is already enormously useful, depending on how much you can speed up those things, and how accurate it can be.

Too slow or too inaccurate and it'll have a strong slowdown factor. But once some threshold been reached, where it makes either of those things faster, I'd probably consider the whole thing "overall useful". Nut of course that isn't the full picture and ignoring all the tradeoffs is kind of cheating, there are more things to consider too as you mention.

I'm guessing we aren't quite over the threshold because it is still very young all things considered, although the ecosystem is already pretty big. I feel like generally things tend to grow beyond their usefulness initially, and we're at that stage right now, and people are shooting it all kind of directions to see what works or not.

Really a lot of innovation, even at the very cutting edge, is about combining old things in new ways, and these are great productivity tools for this.

I've been "vibe coding" quite a bit recently, and it's been going great. I still end up reading all the code and fixing issues by hand occasionally, but it does remove a lot of the grunt work of looking up simple things and typing out obvious code.

It helps me spend more time designing and thinking about how things should work.

It's easily a 2-3x productivity boost versus the old fashioned way of doing things, possibly more when you take into account that I also end up implementing extra bells and whistles that I would otherwise have been too lazy to add, but that come almost for free with LLMs.

I don't think the stereotype of vibe coding, that is of coding without understanding what's going on, actually works though. I've seen the tools get stuck on issues they don't seem to be able to understand fully too often to believe that.

I'm not worried at all that LLMs are going to take software engineering jobs soon. They're really just making engineers more powerful, maybe like going from low level languages to high level compiled ones. I don't think anyone was worried about the efficiency gains from that destroying jobs either.

There's still a lot of domain knowledge that goes into using LLMs for coding effectively. I have some stories on this too but that'll be for another day...

Personally I hope this will materialize, at the very least because there's plenty of discoveries to be made by cross-correlating discoveries already made; the necessary information should be there, but reasoning capability (both that of the model and that added by orchestration) seems to be lacking. I'm not sure if pure chat is the best way to access it, either. We need better, more hands-on tools to explore the latent spaces of LLMs.

In my opinion, there is a very valid argument that the vast majority of things that are patented are not "new" things, because everything builds on something else that came before it.

The things that are seen as "new" are not infrequently something where someone in field A sees something in field B, ponders it for a minute, and goes "hey, if we take that idea from field B, twist it clockwise a bit, and bolt it onto the other thing we already use, it would make our lives easier over in this nasty corner of field A." Congratulations! "New" idea, and the patent lawyers and finance wonks rejoice.

LLMs may not be able to truly "invent" "new" things, depending on where you place those particular goalposts.

However, even a year or two ago - well before Deep Research et al - they could be shockingly useful for drawing connections between disparate fields and applications. I was working through a "try to sort out the design space of a chemical process" type exercise, and decided to ask whichever GPT was available and free at the time about analogous applications and processes in various industries.

After a bit of prodding it made some suggestions that I definitely could have come up on my own if I had the requisite domain knowledge, but would almost certainly never have managed on my own. It also caused me to make a connection between a few things that I don't think I would have stumbled upon otherwise.

I checked with my chemist friends, and they said the resulting ideas were worth testing. After much iteration, one of the suggested compounds/approaches ended up generating the least bad result from that set of experiments.

I've previously sketched out a framework for using these tools (combined with other similar machine learning/AI/simulation tools) to massively improve the energy consumption of industrial chemical processes. It seems to me that that type of application is one where the LLM's environmental cost could be very much offset by the advances it provides.

The social cost is a completely different question though, and I think a very valid one. I also don't think our economic system is structured in such a way that the social costs will ever be mitigated.

Where am I going with this? I'm not sure.

Is there a "ghost in the machine"? I wouldn't place a bet on yes, at least not today. But I think that there is a fair bit of something there. Utility, if nothing else. They seem like a force multiplier to me, and I think that with proper guidance, that force multiplier could be applied to basic research, material science, economics, and "inventions".

Right now, it does seem that it takes someone with a lot of knowledge about the specific area, process, or task to get really good results out of LLMs.

Will that always be true? I don't know. I think there's at least one piece of the puzzle we don't have sorted out yet, and that the utility of the existing models/architectures will ride the s-curve up a bit longer but ultimately flatten out.

I'm also wrong a LOT, so I wouldn't bet a shiny nickel on that.

BTW, the vast majority of JS OAuth libraries are implementing the client side of OAuth. Provider-side implementations are relatively rare, as historically it's mostly only big-name services that ever get to the point of being a OAuth providers, and they tend to build it all in-house and not release code.

…anyway, Gemini pro is a quite good reviewer if you are specific about what you need reviewed and provide relevant dependencies in the context.