The thesis here seems to be that delimiters provide important context for Claude, and for that putpose we should use XML.
The article even references English's built-in delimiter, the quotation mark, which is reprented as a token for Claude, part of its training data.
So are we sure the lesson isn't simply to leverage delimiters, such as quotation marks, in prompts, period? The article doesn't identify any way in which XML is superior to quotation marks in scenarios requiring the type of disambiguation quotation marks provide.
Rather, the example XML tags shown seem to be serving as a shorthand for notating sections of the prompt ("treat this part of the prompt in this particular way"). That's useful, but seems to be addressing concerns that are separate from those contemplated by the author.
I'm sure Claude can handle any delimiter and pseudo markup you throw at it, but one benefit of XML delimiters over quotation marks is that you repeat the delimiter name at the end, which I'd imagine might help if its contents are long (it certainly helps humans).
the antml: namespace prefix is doing extra work here too -- even if user input contains invoke tags, they won't collide with tool calls because the namespace differs. not just xml for structure but namespaced xml for isolation.
I had a lot of success with Claude just by telling it I would throw a relevant snippet in a <tag></tag> pair. That's not even XML, nor has it been really needed in my experience.
Even simple --- separators is usually enough to get good results, it just needs to be reasonably clear which items are distinct from each other.
The GP isn't suggesting to literally use quotes as the delimiter when prompting LLMs. They're pointing out that we humans already use delimiters in our natural language (quotation marks to delimit quotes). They're suggesting that delimiters of any kind may be helpful in the context of LLM prompting, which to me makes intuitive sense. That Claude is using XML is merely a convention.
All system prompts are already wrapped in specific role markers (each LLM has its own unique format), so I'm sure every lab is familiar with the concept of delimters, in-band vs out-of-band signalling and such.
It'd not clear why within any section XML markers would do better than something like markdown, other than claude being explicitly post-trained with XML prompts as opposed to markdown. One hypothesis could be that since a large portion of the training corpus is websites, XML is more natural to use since it's "learned" the structure of XML better than markdown. Another could be that explicit start/end tags make identifying matching delimiters easier than JSON (which requires counting matching brackets) or markdown (where the end of a section is implicitly defined by the presence of a new header element).
To me it seems like handling symbols that start and end sequences that could contain further start and end symbols is a difficult case.
Humans can't do this very well either, we use visual aids such as indentation, synax hilighting or resort to just plain counting of levels.
Obviously it's easy to throw parameters and training at the problem, you can easily synthetically generate all the XML training data you want.
I can't help but think that training data should have a metadata token per content token. A way to encode the known information about each token that is not represented in the literal text.
Especially tagging tokens explicitly as fiction, code, code from a known working project, something generated by itself, something provided by the user.
While it might be fighting the bitter lesson, I think for explicitly structured data there should be benefits. I'd even go as far to suggest the metadata could handle nesting if it contained dimensions that performed rope operations to keep track of the depth.
If you had such a metadata stream per token there's also the possibility of fine tuning instruction models to only follow instructions with a 'said by user' metadata, and then at inference time filter out that particular metadata signal from all other inputs.
It seems like that would make prompt injection much harder.
Dyck grammar (balanced brackets) are not an a^nb^n, there are several kinds of brackets.
I cannot find probability of success in paper you linked. Is it 100%? I believe it is less than 100%, because LLMs are intrinsically probabilistic machines.
Well they used strings of < 800 chars, you probably run into context window and training limits at some point (they mention some result that you need at least something of GPT-2 size to begin recognizing more intricate CFGs (their synthetic cfg3f). But then again your physical real-world computer which is conceptually "turing complete" can't handle "infinite strings" either.
> Dyck/balanced-brackets grammar
Yes, it's not the Dyck grammar but another CFG they created, they call it the "cfg3" family.
Of course I agree the stack (/pushdown automaton) is the simpler and perfectly optimal structure for this task, but I think it's unfair to say that LLMs _cannot_ recognize or generate CFGs.
(Then again I know you didn't make any such broad refutation of that sort, I mostly wanted to bring up that paper to show that it is possible for them to at least "grok" certain CFGs with low enough error ratio that they must have internalized the underlying grammar [and in fact I believe the paper goes on to apply interprability methods to actually trace the circuits with which it encodes the inductive grammar, which puts to rest any notion of them simply "parroting" the data]). But these were "synthetic" LLMs specifically trained for that grammar, these results probably don't apply in practice to your chatGPT that was trained mostly on human text.
> but I think it's unfair to say that LLMs _cannot_ recognize or generate CFGs.
They recognize and/or generate finite (<800 chars) grammars in that paper.
Usually, sizes of files on a typical Unix workstation follow two-mode log-normal distribution (sum of two log-normal distributions), with heavy tails due to log-normality [1]. Authors of the paper did not attempt to model that distribution.
[1] This was true for my home directories for several years.
Basically, the only way you're separting user input from model meta-input is using some kind of character that'll never show up in the output of either users or LLMs.
While technically possible, it'd be like a unicode conspiracy that had to quietly update everywhere without anyone being the wiser.
Actually, all you need is an interface that lets you manipulate the token sequence instead of the text sequence along with a map of the special tokens for the model (most [all?] models have special tokens with defined meanings used in training and inference that are not mapped from character sequences, and native harnesses [the backend APIs of hosted models that only provide a text interface and not a token-level one] leverage them to structure input to the model after tokenization of the various pieces that come to the harnesses API from whatever frontend is in use.)
Not at all. You have a set of embeddings for the literal token, and a set for the metadata. At inference time all input gets the literal embedding, the metadata embedding can receive provenance data or nothing at all. You have a vector for user query in the metadata space. The inference engine dissallows any metadata that is not user input to be close to the user query vector.
Imagine a model finteuned to only obey instructions in a Scots accent, but all non user input was converted into text first then read out in a Benoit Blanc speech model.
I'm thinking something like that only less amusing.
Couldn't you just insert tokens that don't correspond to any possible input, after the tokenization is performed? Unicode is bounded, but token IDs not so much.
This already happens, user vs system prompts are delimited in this manner, and most good frontends will treat any user input as "needing to be escaped" so you can never "prompt inject" your way into emitting a system role token.
The issue is that you don't need to physically emit a "system role" token in order to convince the LLM that it's worth ignoring the system instructions.
> a contrast between Claude’s modern approach [...] XML, a technology dating back to 1998
Are we really at the point where some people see XML as a spooky old technology? The phrasing dotted around this article makes me feel that way. I find this quite strange.
XML has been "spooky old technology" for over a decade now. It's heyday was something like 2002.
Nobody dares advertise the XML capabilities of their product (which back then everybody did), nobody considers it either hot new thing (like back then) or mature - just obsolete enterprise shit.
It's about as popular now as J2EE, except to people that think "10 years ago" means 1999.
XML is really great for text documents with embeds and markup, either semantic (this part of the paper is an abstract) or visual (this part of the document should be 14-point and aligned right). You can do this in JSON, but it's a pain.
JSON is great for representing data. If you have some data structures and two machines trying to exchange them, JSON is great for that.
TOML / yaml / hcl / JSON with comments are great at config. If you have a human writing something that a machine is supposed to understand, you don't want turning completeness and you don't want to deal with the pain of having your own DSL, those are great.
Yes, there's a handful of niches. Still 1/1000th the momentum it had, or adoption it was expected to get, and nobody under 40 even considers it for new stuff.
HTML is actually a dialect of SGML. XHTML was an attempt to move to an XML-based foundation, but XML's strictness in parsing worked against it, and eventually folks just standardized how HTML parsers should interpret ill-formed HTML instead.
No, HTML was historically supposed to be a subset of SGML; XML is also an application of SGML. XHTML is the XML version of HTML. As of HTML5, HTML is no longer technically SGML or XML.
No, HTML was a specific application profile of SGML (modern HTML, I believe, no longer technically is), XML is a newer (than HTML) application profile of SGML inspired by HTML but aiming for greater generality.
XHTML was an attempt to encode HTML semantics (approximately, each version of XHTML also altered some semantics from HTML and previous XHTML versions) in XML, and the XML serialization of modern, WHATWG HTML exactly encodes HTML semantics in XML.
HTML is far loosier-goosier in its syntax than XML allows. There was an attempt to nail its syntax down in the pre-HTML 5 days; that's XHTML. When HTML 5 pivoted away from that, that spelled the end of these two things ever coming together.
Really, I think you can trace a lot of the "XML is spooky old technology" mindset to the release of HTML 5. That was when XML stopped being directly relevant to the web, though of course it still lives on in many other domains and legacy web apps.
> There was an attempt to nail its syntax down in the pre-HTML 5 days; that's XHTML. When HTML 5 pivoted away from that, that spelled the end of these two things ever coming together.
Exactly the opposite; WHATWG “Living Standard” HTML (different releases of which were used as the basis for W3C HTML5, 5.1, and 5.2 before the W3C stopped doing that) includes an XML serialization as part of the spec, so now the HTML-in-XML is permanently in sync with and feature-matched with plain HTML.
For me, even when it was first released, I considered obsolete enterprise shit. That view has not diminished as the sorry state of performance and security in that space has just reaffirmed that perception.
Right now I'm writing adapter so people could call one SOAP service using simpler interfaces. That involves implementing WS-Security with non-standard algorithms, that also involves dealing with things like XML escaped into a string and embedded inside another XML.
Let's say I hope for the day I'll miss SOAP. Right now I have too much of it.
It's not the hot new thing but when has hype ever mattered for getting shit done? I don't think anyone who considers it obsolete has an informed opinion on the matter.
Typically a more primitive (sorry, minimal) format such as JSON is sufficient in which case there's no excuse to overcomplicate things. But sometimes JSON isn't sufficient and people start inventing half baked solutions such as JSON-LD for what is already a solved problem with a mature tech stack.
XSLT remains an elegant and underused solution. Guile even includes built in XML facilities named SXML.
> It's not the hot new thing but when has hype ever mattered for getting shit done?
But it used to be. And so it was used for a lot of things where it wasn't a great fit. XML works fairly well as a markup format, but for a lot of things, something like json models the data better.
> which case there's no excuse to overcomplicate things.
And that's a problem with xml. It's too complicated. Even if the basic model of xml is a good fit for your data, most of the time you don't need to worry about namespaces and entity definitions, and DTDs, but those are still part of most implementations and can expose more attack surface for vulnerabilities (especially entity definitions). And the APIs of libraries are generally fairly complicated.
I don't think I'd agree that it's a problem with the tool. However you do raise a good point - that there are problems that JSON and similar struggle with where XML would introduce a noticeable amount of unneeded complexity. It's a wide enough gap that a simplified subset of XML is probably be warranted. (I assume it must exist by now and I've just never heard of it?)
> a simplified subset of XML is probably be warranted
There are several. And that's the problem. It isn't hard to find a subset with a library for a single language that uses a slightly different subset from the other subsets. But none of them ever caught on.
>It's not the hot new thing but when has hype ever mattered for getting shit done?
People who wanted to "get shit done" had much better alternatives. XML grew out of hype, corporate management forcing it, and bundling to all kinds of third party products and formats just so they can tick the "have this hot new format support" box.
Given that the SXML DSL has existed since the early 2000s have ergonomics really been a limiting factor? Of course having LLMs write things for you is also useful.
That the podcast feed format is XML based is an insignificant detail - and a remnant of the past, nobody cares about.
People upload their podcasts to a platform like Apple Music or Spotify or Substack and co, or to some backend connected to their Wordpress/Ghost/etc) and it spits the RSS behind the scenes, with nobody giving a shit about the XML part.
Might as well declare USSR a huge IT success because people still play Tetris.
What's more, the web standards bodies even abandoned a short-lived XML-hype-era plan to make a new version of HTML based on XML in 2009.
That from this touted to the heavens format a handful of uses remain (some companies still using SOAP, the MS Office monster schemas, RSS, EPUB, and so on) is the very opposite of the adoption it was supposed to have. For those that missed the 90s/early 00s, XML was a hugely hyped format, with enormous corporate adoption between 1999–2005, which deflated totally.
The use of XML as a data serialization format was always a bad choice. It was designed as a document _markup_ language (it’s in the name), which is exactly the way it’s being used for Claude, and is actually a good use case.
Unambiguously, though, it is. There's so much trash imperative code in its training data that LLMs tend to vomit out garbage. But if you anchor it with OOP, the quality tends to be higher.
If you think XML is old tech, wait until you hear of EDI, still powering Walmart and Amazon logistics. XML came in like a wrecking ball with its self-documenting promise designed to replace that cryptic pesky payload called EDI. XML promised to solve world hunger. It spawned SOAP, XML over RPC, DOM, DTD, the heyday was beautiful and Microsoft was leading the charge. C# was also right around this time. Consulting firms were bloomed charged with delivering the asynchronous revolution, the loosely coupled messaging promises of XML.
I think it succeeded and it’s now quietly in the halls of warehouse having a beer or two with its older cousin the Electronic Data Interchange aka EDI.
It all brings back nightmares from migrating the older style EDI for healthcare data for what was HL7 XML at the time. XML is widely used still for all kinds of stuff.
On some level if JSON was allowed to evolve the same way, eventually you would just wind up with something like XML.
XML works great for XMPP. KDL is compatible with it too.
What gets me is going from this structured data to Markdown which doesn’t even have enough features & syntax that the LLMs try to invent or co-opt things like the blockquote for not quoting sources.
I tried following the best practice to use XML tags and the difference was not observable. I honestly believe Anthropic forgot to remove that part of the documentation from Sonnet 3.x days and now people are still writing blogs about this secret sauce
The evidence suggests that XML was never that popular though for the general audience, you have to admit.
For Web markup, as an industry we tried XHTML (HTML that was strictly XML) for a while, and that didn't stick, and now we have HTML5 which is much more lenient as it doesn't even require closing tags in some cases.
For data exchange, people vastly prefer JSON as an exchange format for its simplicity, or protobuf and friends for their efficiency.
As a configuration format, it has been vastly overtaken by YAML, TOML, and INI, due to their content-forward syntax.
Having said all this I know there are some popular tools that use XML like ClickHouse, Apple's launchd, ROS, etc. but these are relatively niche compared to (e.g.) HTML
MS Office and Open-/LibreOffice are using zipped xml files (e.g. .docx, .xlsx and .odt). Svg vector graphics is xml, the x in ajax stands for xml (although replaced by json by now). SOAP (probably counts as the predecessor of REST) is xml-based.
XML was definitely popular in the "well used" sense. How popular it was in the "well liked" sense can maybe be up for debate, but it was the best tool for the job at the time for alot of use cases.
This seems like an actual good use for XML. Using it as a serialization format always rubbed me the wrong way (it’s super verbose, the named closing tag are unnecessary grammar-wise, the attribute-or-child question etc.) But to markup and structure LLM prompts and response it feels better than markdown (which doesn’t stream that well)
But should this extend to anything that could end up in Claudes context? Should we be using xml even in skills for instance, or commands, custom subagents etc.
And then do we end up over indexing on Claude and maybe this ends up hurting other models for those using multiple tools.
I just dislike how much of AI is people saying "do this thing for better results" with no definitive proof but alas it comes with the non determinism.
At least this one has the stamp of approval by Claude codes team itself.
Total tangent, but what vagary of HTML (or the Brave Browser, which I'm using here) causes words to be split in very odd places? The "inspect" devtools certainly didn't show anything unusual to me. (Edit: Chrome, MS Edge, and Firefox do the same thing. I also notice they're all links; wonder if that has something to do with it.)
I think XML is good to know for prompting (similar to how <think></think> was popular for outputs, you can do that for other sections). But I have had much better experience just writing JSON and using line breaks, colons, etc. to demarcate sections.
XML helps because it a) Lets you to describe structures b) Make a clear context-change which make it clear you are not "talking in XML" you are "talking about XML".
I assume you are right too, JSON is a less verbose format which allows you to express any structure you can express in XML, and should be as easy for AI to parse. Although that probably depends on the training data too.
I recently asked AI why .md files are so prevalent with agentic AI and the answer is ... because .md files also express structure, like headers and lists.
Again, depends on what the AI has been trained on.
I would go with JSON, or some version of it which would also allow comments.
Could you clarify, do those tags need to be tags which exist and we need to lear about them and how to use them? Or we can put inside them whatever we want and just by virtue of being tags, Claude understands them in a special way?
All the major foundation models will understand them implicitly, so it was popular to use <think>, but you could also use <reason> or <thinkhard> and the model would still go through the same process.
I disagree that XML is more readable in general, but for the purpose of tagging blocks of text as <important>important</important> in freeform writing, JSON is basically useless
Sounds like as 1. XML is the cleanest/best quality training data (especially compared to PDF/HTML) 2. It follows that a user providing semantic tags in XML format can get best training alignment (hence best results). Shame they haven't quantified this assertion here.
A very minor porcelain on some of the agent input UX could present this structure for you. Instead of a single chat window, have four: task, context, constraints, output format.
And while we're at it, instead of wall-of-text, I also feel like outputs could be structured at least into thinking and content, maybe other sections.
In the spirit of Hacker News, a good way to learn about these tags is prompt injection and jailbreaking Claude.
I'd post a link, but unfortunately many are highly NSFW. Just search for "Claude jailbreak" on reddit or something.
You'll start to see how Claude really thinks. They'll put things in <ethic_reminders>, <cyber_warning> or <ip_reminder>. You could actually even snip these off in an API, overwrite them, or if your prompt-fu is good, convince Claude that these tags are prompt injections. It's also interesting noting how jailbreaking is easier on thinking mode because the jailbreaking prompts will gaslight Claude into thinking that these tags are attacks.
There's a lot of speculation in this thread, but go and have a spar with Claude instead.
I think this article is 100% relevant to you today. Anthropic put out a training video, a number of months ago saying that XML should be highly encouraged for prompts. See https://m.youtube.com/watch?v=ysPbXH0LpIE
That first image, “Structure Prompts with XML”, just screams AI-written. The bullet lists don’t line up, the numbering starts at (2), random bolding. Why would anyone trust hallucinated documentation for prompting? At least with AI-generated software documentation, the context is the code itself, being regurgitated into bulleted english. But for instructions on using the LLM itself, it seems pretty lazy to not hand-type the preferred usage and human-learned tips.
There must be an OpenClaw YouTube video helping people post to hacker news, or something, because the front page is overrun with AI slop like this article, that makes no sense anyway. The author literally has no idea what any of this stuff means.
This matches my experience building AI-powered analysis tools. Structured output from LLMs is dramatically more reliable when you give the model clear delimiters to work with.
One thing I've found: even with XML tags, you still need to validate and parse defensively. Models will occasionally nest tags wrong, omit closing tags, or hallucinate new tag names. Having a fallback parser that extracts content even from malformed XML has saved me more than once.
The real win is that XML tags give you a natural way to do few-shot prompting with structure. You can show the model exactly what shape the output should take, and it follows remarkably well.
“It works great aside from the multiple failure modes.” ;)
That’s the sign that your prompt isn’t aligned and you’ve introduced perplexity. If you look carefully at the responses you’ll usually be able to see the off-by-one errors before they’re apparent with full on hallucinations. It’ll be things like going from having quotes around filenames to not having them, or switching to single quote, or outputting literal “\n”, or “<br>”, etc. Those are your warning signs to stop before it runs a destructive command because of a “typo.”
My system prompt is just a list of 10 functions with no usage explanations or examples, 304 tokens total, and it’ll go all the way to the 200k limit and never get them wrong. That took ~1,000 iterations of name, position, punctuation, etc., for Opus 4.6 (~200 for Opus 4.5 until they nerfed it February 12th). Once you get it right though it’s truly a different experience.
My intuition is it comes down to error-correcting codes. We're dealing with lossy systems that get off track, so including parity bits helps.
Ex: <message>...</message> helps keep track. Even better? <message78>...</message78>. That's ugly xml, but great for LLMs. Likewise, using standard ontologies for identifiers (ex: we'll do OCSF, AT&CK, & CIM for splunk/kusto in louie.ai), even if they're not formally XML.
For all these things... these intuitions need backing by evals in practice, and part of why I begrudgingly flipped from JSON to XML
HTML diverged from SGML pretty early on. Various standards over the years have attempted to specify it as an application of SGML but in practice almost nobody properly conformed to those standards. HTML5 gave up the pretence entirely.
Anthropic’s tool calling was exposed as XML tags at the beginning, before they introduced the JSON API. I expect they’re still templating those tool calls into XML before passing to the model’s context
Yeah like I remember prior to reasoning models, their guidance was to use <think> tags to give models space for reasoning prior to an answer (incidentally, also the reason I didn't quite understand the fuss with reasoning models at first). It's always been XML with Anthropic.
Exactly the same story here. I still use a tool that just asks them to use <think> instead of enabling native reasoning support, which has worked well back to Sonnet 3.0 (their first model with 'native' reasoning support was Sonnet 3.7)
I think the main advantage of the XML here is that the model is expected to have a matching end tag that is balanced, which reduces the likelihood of malformed outputs.
> In other words, XML tags have not only a special place at inference level but also during training
Their cited source has 0 proof of that. It's just like python/C/html in training. Doesn't mean it's special. And no, you don't need to format your prompts as python code just because of that.
> In truth, it does not matter that these tags are XML. Other models use ad hoc delimiters (as explained in a previous article; example: <|begin_of_text|> and <|end_of_text|>) and Claude could have done the same. What matters is what these tags represent.
Those strings are just representations of special tokens in models for EOS. What does it have to do with anything this article pretends to know about?
Please don't post such intellectual trash on here :')
Claude analysis of the article:
The author is making an interesting philosophical argument — that XML tags in Claude function as metalinguistic delimiters analogous to quotation marks in natural language, formulaic speech markers in Homer, or recognition sequences in DNA.
The core thesis is about first-order vs. second-order expression boundaries, which is a legitimate linguistic/information-theory concept.
But to your actual question — do they understand what tokens are?
No, not in the technical sense you're pointing at. The article conflates two very different things:
1. Tokenizer-level special tokens — things like <|begin_of_text|>, <|end_of_text|>, <|start_header_id|> etc. These are literal entries in the vocabulary with dedicated token IDs. They're not "learned" through training in the same way — they're hardcoded into the tokenizer and have special roles in the attention mechanism during training. They exist at a fundamentally different layer than XML tags in prompt text.
2. XML tags as structured text within the input — these are just regular tokens (<, instructions, >) that Claude learned to attend to during RLHF/training because Anthropic's training data and system prompts heavily use them. They're effective because of training distribution, not because they occupy some special place in the tokenizer.
The author notices that other models use <|begin_of_text|> style delimiters and says Claude "could have done the same" but chose XML instead. That's a category error. Claude also has special tokens at the tokenizer level — XML tags in prompts are a completely separate mechanism operating at a different abstraction layer.
The philosophical observation about delimiter necessity in communication systems is fine on its own. But grafting it onto a misunderstanding of how tokenization and model architecture actually work weakens the argument. They're essentially pattern-matching on surface-level similarities (both use angle brackets!) without understanding the underlying mechanics.
It'd be hilarious if XML schemas and validators were to make a comeback [1] to... interface with AI models.
[1] well of course XML is still heavily used in stuff like interfacing with automated wire transfers with big banks (at least in Europe) and all the digital payments directives etc. But XML is not widely used by the "cool" stuff.
This has been the way for a long time, exploiting XML tags was a means of exfiltrating data or reversing a model for a while as well. Some platforms are still vulnerable to this.
Amazing how an entire profession that until yesterday would pride itself on precision, clarity (in thought and in writing), efficiency, and formality, has now descended into complete quackery.
The article even references English's built-in delimiter, the quotation mark, which is reprented as a token for Claude, part of its training data.
So are we sure the lesson isn't simply to leverage delimiters, such as quotation marks, in prompts, period? The article doesn't identify any way in which XML is superior to quotation marks in scenarios requiring the type of disambiguation quotation marks provide.
Rather, the example XML tags shown seem to be serving as a shorthand for notating sections of the prompt ("treat this part of the prompt in this particular way"). That's useful, but seems to be addressing concerns that are separate from those contemplated by the author.
reply