Mostly read like a normal article if you skip over the parts about using agents, which I did.
There could not possibly be a single thing in the world more boring than listening to someone describe using an AI agent. Might as well describe in arduous detail how you use a gas pump or a grocery store checkout.
I found the bit about using an agent to produce a minimal repro case interesting.
I'm very much on the side of "artisanally writing all code by hand" in terms of preference, but I have to admit that sometimes it puts me in a bind where there's stuff I wish I could do but don't because it's too time-consuming for the value provided. Isolating repro cases sometimes falls into that bucket.
This seems like a good use case for AI. Even if the resulting code isn't great, as long as it repros the issue and is fairly small, it's fine. It's eventually throwaway code anyway. At best, it will be harvested for a test case, but otherwise it's gone once the bug is fixed.
I think the problem is that the public is often being encouraged to use AI in ways that are not productive, or are misleading... At least at this point in time.
The part where Claude specifically tipped them off felt helpful to include. Stating that you used Claude to do a first pass just sounds like "I opened Vscode with highlighting to do a first pass" and that doesn't sound so relevant.
I might be too used to using coding agents in various parts of my workflow, and others are still getting acquainted, or others find it still much different than just another standard debugging tool.
And fwiw it's probably not Claude's fault that emoji fonts load this slow, though. Wtf Safari?
Maybe off topic, but I couldn't help thinking that "we need to show a heart icon" -> "let's use a heart emoji because it's easy" -> "let's use a specific emoji font for consistency across platforms" -> "let's import it from Google Fonts every time" seems like a problematic developer mindset.
A better heuristic is always keep in mind not only developer efficiency, but also program efficiency. I'm not saying optimize everything, but keep program efficiency in mind at least a little bit. In this case, that would've led the developer to download a tiny SVG or PNG and serve it from the app itself. Which would've avoided the problem in the post, and maybe other problems as well.
While in this case we’d included the emoji font for displaying user content in another part of the app, the hazard of letting a “simple” approach expand and get out of hand is part of what I wanted to convey in writing this.
Not OT at all. Emojis everywere are ridiculous. And coding agents love them! They put emojis in Python log lines which inevitably break the console, and of course in web pages. Logs don't need emojis. Not sure if anything does.
I love using emojis in my log lines, especially symbols for info/warn/error, but it does add another layer of complexity as you have to go through so many things to make sure the text is now rendered in the right font, has Unicode support enabled, etc, etc.
I agree that the font and emoji hops aren’t great for complexity or performance, but the problem in the post was in the rendering of a tiny SVG; serving it directly would not have avoided the problem.
> And despite being the least-bad approach for web frontends today, the React ecosystem...
As if anyone has seriously tried anything other than the "reactive UI hacked together with callbacks or proxies, with weird XML-like syntax in the JS code" paradigm for the last 10 years.
At this point I just have to conclude that anyone who believes this stuff is good is either too indoctrinated into this workflow or just lacks ability to do even the tiniest amount of cost/benefit analysis.
I'd give people the benefit of the doubt. Personally, having built UI with Win32, WinForms, VisualBasic, Cocoa/Interface Builder, Qt, Tcl/Tk, XSLT, vanilla HTML/JS, jQuery, Backbone, Ember, Knockout, Bootstrap, MooTools, YUI, ExtJS, Svelte, Web Components, and React (including Preact, SolidJS…)… I'll happily choose the React approach. The only other one I would even describe as "good" was Qt.
I also don't get why "XML-like syntax in the JS code" is even a point worth complaining about. If anything, we should be encouraging people to experiment with more DSLs for building UIs. Make your tools suit the task. Who the fuck cares, if it's clear and productive?
But it is not really XML like syntax, is it? It is still a string, even if a template string or whatever it is called, no?
That still leaves the door open for XSS. A good (proper?) (e?)DSL would have the things that make the DOM as keywords in the language, and then we could ensure, that things which should merely be a text, are really only rendered as text, not injected DOM nodes. And the next failure is, that this DSL that is jsx needs to rename HTML attributes, because of overlap with JS keywords like "class". It lacks the awareness of context and therefore is not truly like HTML, no matter how hard it tries to be. It also comes with hacks like "<>" or fragment.
Overall it is usable, but not a particularly well made DSL. It might be as good as it gets with JS.
For inspiration check SXML in various lisps, which comes with immunity to XSS and which works just like the rest of the language, can be structurally pattern matched upon and iterated through, like a proper tree structure.
> It is still a string, even if a template string or whatever it is called, no?
No.
> That still leaves the door open for XSS.
The door for that in React is called `dangerouslySetInnerHTML`, but it's extremely rarely used.
> jsx needs to rename HTML attributes, because of overlap with JS keywords like "class"
That's not really inherent to JSX, just React's use of it. SolidJS, for example, uses `class` instead. But in any case – JSX didn't make up those names. Those are the property names on JavaScript's DOM classes. The fact that there's confusion between "attributes" and "properties" is pretty baked-in to the Web platform, even causing confusion in standard Web Components. Every DOM library and framework (even jQuery) has needed to decide whether it's operating on properties or attributes.
const div = document.createElement('div');
div.className = 'foo';
> It also comes with hacks like "<>" or fragment.
The DOM has the same concept, DocumentFragment. How else would you represent e.g. "two sibling nodes with no parent node"?
> It lacks the awareness of context and therefore is not truly like HTML.
On the contrary, I'd argue it has way more context. It knows, and will warn you, if you try to do any DOM element nesting that the HTML spec forbids, for example.
> can be structurally pattern matched upon and iterated through, like a proper tree structure.
You are literally describing the output of JSX. Glad you like it ;)
Reactive UIs with unidirectional data binding (very important) seem to be the sweet spot. Spreadshees, which pioneered it in consumer software, still reign supreme.
React is quite fast, and is very compact (preact is half the sizef htmx). It seems to be the sweet spot for making rich web UIs.
In the end, this all was a red herring. The problem was in CoreSVG taking 1400ms to render an emoji, clearly a regression. A tweet would suffice to communicate this nugget, but for some reason the author wrote a long and winding piece.
Did the author start a client-side debugging process without running any kind of diagnostics in the dev tools?
To me this sounds like "I saw some slowness in my application, so naturally I started adjusting indexes in my database." Without doing any upfront research, you're basically just throwing darts and hoping one sticks. This type of approach to debugging has never served me well.
> At that point, I reached for an age-old tool that has gotten more useful in the modern age: binary search. That is, you explain the symptom to your coding agent. Then you have it repeatedly remove stuff from your code that might be causing the problem
Can someone give me some high level pointers on how to setup this scaffolding?
When I read the first sentence, I expected the author to use `git bisect`.
However, what the author seems to have done is used a prompt with claude that probably looked something like this:
"Some piece of code is causing the page to load very slowly. To debug this, I'd like to use binary search, where we keep commenting/uncommenting 50% of the remaining code, and then I manually check if the page is still very slow. Let's start now; Comment out a component (or parts of a component) that you estimate is 50% of the page, and I will tell you if the page is still slow."
I do this all the time in a dumb but effective way. Add logging statements to code paths that drop timing info. Another dumb but effective way, instead of using a step through debugger, is drop "here, value is {val}". Telling claude to do this is trivial, it's quick, and it can read its own output and self-solve the problem all with just the code itself.
IMHO git bisect is slower, especially depending on the reload/hot-reload/compile/whatever process your actual app is using.
How well agents can do this is mostly proportional to how well they can understand and navigate your codebase broadly.
There are various contributing factors to this, but they include clear docs, notes and refactors that clear up parts the agent commonly gets confused by, choosing boring technology (your dependencies are well understood) and access to command-line tools that let it lint + typecheck + test the code. A lot of the scaffolding and wiring necessary are built into Cursor and Claude Code themselves now. Hope that helps!
Enjoyed this piece! I use binary search to investigate data pipeline issues. Didn’t think to use it to debug features with agents which was a very cool approach.
In this case, I had made an overlarge squashed merge that included both the Intercom integration (a suspiciously likely cause of slowness) and the feedback button that added the heart – so I needed to go deeper to figure out the true cause. (Noto Emoji was in the app from before, but wasn't triggered in the dashboard until we added an emoji there.)
This is a legitimately fun piece about a bug (or extraordinary levels of inefficiency) in CoreSVG, manifested in massive computational loads to display a single SVG fallback for a colour-specified emoji.
But, isn't the heart emoji red anyways, across basically every font that has emojis? I mean, even with variations. I'm not sure what COLRv1 brings to that table for that scenario. Although maybe the special font is overkill if you really wanted to do something crazy with an emoji or text, and it seems to focus on gradients and the like.
Maybe this is why they humorously blame Claude for getting them to use that font and its affordances in the first place.
I don't think the blog post itself is using that emoji font. The screenshot on the Noto Emoji Github page[0] doesn't look like it's using any gradients for the heart emoji, just flat shading. But it is using gradients for some of the other emojis (e.g. the croissant), and obviously the SVG fallback is all or nothing, not per-glyph.
You need to look closer; the heart emoji has a flat fill, but a gradient in its outline stroke, from lighter-than-red near the top, to darker-than-red on the bottom.
As a bit more on this, until this piece I was oblivious to this COLRv1 thing, which is adding more of SVG-style functionality to already vector font standards.
The conclusion of "yes, Claude helped fix this, but it also caused it by recommending an emoji font" seems a bit disingenuous to me. Using an emoji font is a good suggestion, it's not like Claude (or anyone) could have known there's an SVG but that will cause this slowness.
Still, I'm going to quote the living daylights out of "These coding agents are very much like a power saw. Profoundly useful, and proportionately dangerous."
Even though this particular case may not be fair, the comparison feels like a very fair one to me. The notion that these things can be very valuable in capable hands, but costly in others.
FTFY. I have created a font with a 1BP LLM inside which executes each time the glyph is rendered and figures out the appropriate representation. It will also recursively check the contents of the page and punch up the text.
The LLM contains copies of QEMU, so it can bootstrap itself throughout the enterprise. Naturally, if it finds another LLM, it replaces it. Running an OS and LLM is redundant, so eventually all the machines boot directly into the LLM. It can emulate all the popular desktops, so users won't notice the difference. ;-)
Mostly read like a normal article if you skip over the parts about using agents, which I did.
There could not possibly be a single thing in the world more boring than listening to someone describe using an AI agent. Might as well describe in arduous detail how you use a gas pump or a grocery store checkout.
I found the bit about using an agent to produce a minimal repro case interesting.
I'm very much on the side of "artisanally writing all code by hand" in terms of preference, but I have to admit that sometimes it puts me in a bind where there's stuff I wish I could do but don't because it's too time-consuming for the value provided. Isolating repro cases sometimes falls into that bucket.
This seems like a good use case for AI. Even if the resulting code isn't great, as long as it repros the issue and is fairly small, it's fine. It's eventually throwaway code anyway. At best, it will be harvested for a test case, but otherwise it's gone once the bug is fixed.
Maybe consider that other people are not as au fait with AI.
There are plenty of people learning to use AI and this article helps them.
A reasonable amount of repetition of "things you should know" is good.
The article is well written because you could skip the parts you knew, and learn from the parts you were unfamiliar with.
I think the problem is that the public is often being encouraged to use AI in ways that are not productive, or are misleading... At least at this point in time.
The part where Claude specifically tipped them off felt helpful to include. Stating that you used Claude to do a first pass just sounds like "I opened Vscode with highlighting to do a first pass" and that doesn't sound so relevant.
I might be too used to using coding agents in various parts of my workflow, and others are still getting acquainted, or others find it still much different than just another standard debugging tool.
And fwiw it's probably not Claude's fault that emoji fonts load this slow, though. Wtf Safari?
Maybe off topic, but I couldn't help thinking that "we need to show a heart icon" -> "let's use a heart emoji because it's easy" -> "let's use a specific emoji font for consistency across platforms" -> "let's import it from Google Fonts every time" seems like a problematic developer mindset.
A better heuristic is always keep in mind not only developer efficiency, but also program efficiency. I'm not saying optimize everything, but keep program efficiency in mind at least a little bit. In this case, that would've led the developer to download a tiny SVG or PNG and serve it from the app itself. Which would've avoided the problem in the post, and maybe other problems as well.
Not off topic at all!
While in this case we’d included the emoji font for displaying user content in another part of the app, the hazard of letting a “simple” approach expand and get out of hand is part of what I wanted to convey in writing this.
Not OT at all. Emojis everywere are ridiculous. And coding agents love them! They put emojis in Python log lines which inevitably break the console, and of course in web pages. Logs don't need emojis. Not sure if anything does.
I love using emojis in my log lines, especially symbols for info/warn/error, but it does add another layer of complexity as you have to go through so many things to make sure the text is now rendered in the right font, has Unicode support enabled, etc, etc.
I agree that the font and emoji hops aren’t great for complexity or performance, but the problem in the post was in the rendering of a tiny SVG; serving it directly would not have avoided the problem.
I would say just reusing widely-used emojis you have already downloaded would be less error prone
... assuming it all works ofc (though you could say that about serving svgs too)
> And despite being the least-bad approach for web frontends today, the React ecosystem...
As if anyone has seriously tried anything other than the "reactive UI hacked together with callbacks or proxies, with weird XML-like syntax in the JS code" paradigm for the last 10 years.
At this point I just have to conclude that anyone who believes this stuff is good is either too indoctrinated into this workflow or just lacks ability to do even the tiniest amount of cost/benefit analysis.
I'd give people the benefit of the doubt. Personally, having built UI with Win32, WinForms, VisualBasic, Cocoa/Interface Builder, Qt, Tcl/Tk, XSLT, vanilla HTML/JS, jQuery, Backbone, Ember, Knockout, Bootstrap, MooTools, YUI, ExtJS, Svelte, Web Components, and React (including Preact, SolidJS…)… I'll happily choose the React approach. The only other one I would even describe as "good" was Qt.
I also don't get why "XML-like syntax in the JS code" is even a point worth complaining about. If anything, we should be encouraging people to experiment with more DSLs for building UIs. Make your tools suit the task. Who the fuck cares, if it's clear and productive?
But it is not really XML like syntax, is it? It is still a string, even if a template string or whatever it is called, no?
That still leaves the door open for XSS. A good (proper?) (e?)DSL would have the things that make the DOM as keywords in the language, and then we could ensure, that things which should merely be a text, are really only rendered as text, not injected DOM nodes. And the next failure is, that this DSL that is jsx needs to rename HTML attributes, because of overlap with JS keywords like "class". It lacks the awareness of context and therefore is not truly like HTML, no matter how hard it tries to be. It also comes with hacks like "<>" or fragment.
Overall it is usable, but not a particularly well made DSL. It might be as good as it gets with JS.
For inspiration check SXML in various lisps, which comes with immunity to XSS and which works just like the rest of the language, can be structurally pattern matched upon and iterated through, like a proper tree structure.
> It is still a string, even if a template string or whatever it is called, no?
No.
> That still leaves the door open for XSS.
The door for that in React is called `dangerouslySetInnerHTML`, but it's extremely rarely used.
> jsx needs to rename HTML attributes, because of overlap with JS keywords like "class"
That's not really inherent to JSX, just React's use of it. SolidJS, for example, uses `class` instead. But in any case – JSX didn't make up those names. Those are the property names on JavaScript's DOM classes. The fact that there's confusion between "attributes" and "properties" is pretty baked-in to the Web platform, even causing confusion in standard Web Components. Every DOM library and framework (even jQuery) has needed to decide whether it's operating on properties or attributes.
> It also comes with hacks like "<>" or fragment.The DOM has the same concept, DocumentFragment. How else would you represent e.g. "two sibling nodes with no parent node"?
> It lacks the awareness of context and therefore is not truly like HTML.
On the contrary, I'd argue it has way more context. It knows, and will warn you, if you try to do any DOM element nesting that the HTML spec forbids, for example.
> can be structurally pattern matched upon and iterated through, like a proper tree structure.
You are literally describing the output of JSX. Glad you like it ;)
https://github.com/jquery/jquery/blob/df16f7360d4db41de64589...
Whoa backbone, ember, knockout, what throwbacks right there.
Reactive UIs with unidirectional data binding (very important) seem to be the sweet spot. Spreadshees, which pioneered it in consumer software, still reign supreme.
React is quite fast, and is very compact (preact is half the sizef htmx). It seems to be the sweet spot for making rich web UIs.
In the end, this all was a red herring. The problem was in CoreSVG taking 1400ms to render an emoji, clearly a regression. A tweet would suffice to communicate this nugget, but for some reason the author wrote a long and winding piece.
> Noto Color Emoji is a Google font that is helpful in that it gives you consistent emoji rendering across platforms.
I’m not sure how that is helpful if users are used to the emoji look of their respective platform.
> Naturally, I blamed React.
Did the author start a client-side debugging process without running any kind of diagnostics in the dev tools?
To me this sounds like "I saw some slowness in my application, so naturally I started adjusting indexes in my database." Without doing any upfront research, you're basically just throwing darts and hoping one sticks. This type of approach to debugging has never served me well.
> At that point, I reached for an age-old tool that has gotten more useful in the modern age: binary search. That is, you explain the symptom to your coding agent. Then you have it repeatedly remove stuff from your code that might be causing the problem
Can someone give me some high level pointers on how to setup this scaffolding?
When I read the first sentence, I expected the author to use `git bisect`.
However, what the author seems to have done is used a prompt with claude that probably looked something like this:
"Some piece of code is causing the page to load very slowly. To debug this, I'd like to use binary search, where we keep commenting/uncommenting 50% of the remaining code, and then I manually check if the page is still very slow. Let's start now; Comment out a component (or parts of a component) that you estimate is 50% of the page, and I will tell you if the page is still slow."
`git bisect` is interesting option. I haven't heard about it before. Thanks for info. Still learning something ;)
I'm old school. I used to do "manual bisection" on git history by just `git checkout <commit_id>` until I find first introducing bug commit.
Then another "bisection" on commit changes until minimal change found.
Deterministic bugs are quite "fine". For me personally worst are randomly occurring bugs in specific conditions for eg. some race conditions.
I do this all the time in a dumb but effective way. Add logging statements to code paths that drop timing info. Another dumb but effective way, instead of using a step through debugger, is drop "here, value is {val}". Telling claude to do this is trivial, it's quick, and it can read its own output and self-solve the problem all with just the code itself.
IMHO git bisect is slower, especially depending on the reload/hot-reload/compile/whatever process your actual app is using.
How well agents can do this is mostly proportional to how well they can understand and navigate your codebase broadly.
There are various contributing factors to this, but they include clear docs, notes and refactors that clear up parts the agent commonly gets confused by, choosing boring technology (your dependencies are well understood) and access to command-line tools that let it lint + typecheck + test the code. A lot of the scaffolding and wiring necessary are built into Cursor and Claude Code themselves now. Hope that helps!
Not sure what you mean, just have a coding agent (e.g. Claude Code) and talk to it.
"git bisect" usually does the trick.
Enjoyed this piece! I use binary search to investigate data pipeline issues. Didn’t think to use it to debug features with agents which was a very cool approach.
I think you could have discovered this bug more easily by looking at the commit(s) that were made when the problem started.
This is a great technique!
In this case, I had made an overlarge squashed merge that included both the Intercom integration (a suspiciously likely cause of slowness) and the feedback button that added the heart – so I needed to go deeper to figure out the true cause. (Noto Emoji was in the app from before, but wasn't triggered in the dashboard until we added an emoji there.)
This is a legitimately fun piece about a bug (or extraordinary levels of inefficiency) in CoreSVG, manifested in massive computational loads to display a single SVG fallback for a colour-specified emoji.
But, isn't the heart emoji red anyways, across basically every font that has emojis? I mean, even with variations. I'm not sure what COLRv1 brings to that table for that scenario. Although maybe the special font is overkill if you really wanted to do something crazy with an emoji or text, and it seems to focus on gradients and the like.
Maybe this is why they humorously blame Claude for getting them to use that font and its affordances in the first place.
It's not solid red. It has shading.
I don't think the blog post itself is using that emoji font. The screenshot on the Noto Emoji Github page[0] doesn't look like it's using any gradients for the heart emoji, just flat shading. But it is using gradients for some of the other emojis (e.g. the croissant), and obviously the SVG fallback is all or nothing, not per-glyph.
[0] https://github.com/googlefonts/noto-emoji
You need to look closer; the heart emoji has a flat fill, but a gradient in its outline stroke, from lighter-than-red near the top, to darker-than-red on the bottom.
As a bit more on this, until this piece I was oblivious to this COLRv1 thing, which is adding more of SVG-style functionality to already vector font standards.
https://nabla.typearture.com/
https://developer.chrome.com/blog/colrv1-fonts
My natural cynicism is to ask "should a font really do this?" But I guess it's pretty neat.
The conclusion of "yes, Claude helped fix this, but it also caused it by recommending an emoji font" seems a bit disingenuous to me. Using an emoji font is a good suggestion, it's not like Claude (or anyone) could have known there's an SVG but that will cause this slowness.
Still, I'm going to quote the living daylights out of "These coding agents are very much like a power saw. Profoundly useful, and proportionately dangerous."
Even though this particular case may not be fair, the comparison feels like a very fair one to me. The notion that these things can be very valuable in capable hands, but costly in others.
... and a broken world.
How infuriating it is to see complexity so spuriously piled up upon an already holy mess.
FTFY. I have created a font with a 1BP LLM inside which executes each time the glyph is rendered and figures out the appropriate representation. It will also recursively check the contents of the page and punch up the text.
The LLM contains copies of QEMU, so it can bootstrap itself throughout the enterprise. Naturally, if it finds another LLM, it replaces it. Running an OS and LLM is redundant, so eventually all the machines boot directly into the LLM. It can emulate all the popular desktops, so users won't notice the difference. ;-)