Pretty nice. I've been using LLMs to generate different Python and JS tools for wrangling data for ontology engineering purposes.
More recently, I've found a lot of benefit from using the extended thinking mode in GPT-5 and -5.1. It tends to provide a fully functional and complete result from a zero-shot prompt. It's as close as I've gotten to pair programming with a (significantly) more experienced coder.
One functional example of that (with 30-50% of my own coding, reprompting and reviews) is my OntoGSN [1] research prototype. After a couple of weeks of work, it can handle different integration, reasoning and extension needs of people working in assurance, at least based on how I understood them. It's an example of a human-AI collab that I'm particularly proud of.
Nice, I do this often enough that I created a bookmarklet to download an HTML file from clipboard after copying ChatGPT's code block.
I've also been using LLMs to create and maintain a "work assist" Chrome extension that I load unpacked from a local directory. Whenever I notice a minor pain point, I get the LLM to quickly implement a remedy. For example, I usually have several browser tabs open for Jira, and they all have the same company logo as the favicon, so my Chrome extension changes the favicon to be the issue type icon (e.g. Bug, Story, etc) when the page loads. It saves a little time when I'm looking for a specific ticket I've already opened.
I also find that sticking to a single file makes coding agents perform better (fewer surgical edits, faster outputs, sensible changes, etc).
Not sure why, but the moment the file is split into files and subfolders, coding agents tend to do a lot more changes that what is absolutely necessary. That way a single html file wins!
My tool collection [0] is inspired by yours, with a handful of differences. I'm only at 53 tools at the moment.
What I did differently:
Hosted on Cloudflare Pages. This gives you preview URLs for pull requests out the box. This might be possible with Github Pages but I haven't checked. I've used Vercel for similar projects in the past. Cloudflare seems to have the odd failed build that needs a kick from their dashboard.
Some tools can make use of Workers/Functions for backend processing and secrets. I try to keep these to a minimum but they're occasionally useful.
I have an AGENTS.md that's updated with a Github action to automatically pull in Claude-style Skills from the .skills directory. I blogged about this pattern and am still waiting for a standard to evolve [2].
I have a base stylesheet that I instruct agents to pull in. This gives a bit of consistency and also let's them use Tailwind, which they'd seem to love.
Thanks for showing this! It’s cool, and I enjoyed reading through some of the code.
Note that I tried to use some of the regex tools that needed LLMs and got a rate limit error.
These are great. Something you might find interesting is that you can expose a google sheet to have an interactive database. I have a map similar to yours, but with surf spots. Maybe defeats the point, but I find it handy
Edit: come to think of it, I should revisit it now that everyone can vibe code. The sheet was to allow people to add to it, now maybe easier for me to take a message and ask an agent to update the html directly
Couple of unsolicited comments: first is that on mobile, the featured badge sits on top of the right facing arrow. Second is that the bubble level seems to be upside down? The bubble sinks rather than floats at least on my pixel
Really enjoyed this, interesting read as always! It reminded me of Google Labs’ recent GenTabs project [1], and also of a recent ACM paper on user-assembled LLM-mediated tools from web content [2]. Feels like similar concepts are emerging in multiple places, all centered around lightweight intent-driven tools rather than traditional apps, which I think makes a lot of sense. Curious how this will evolve!
Great timing - I've taken up vibecoding and picked personal tools with simple HTML/CSS/JS/python stack as the learning ground too.
Personal tools seem like a reasonable place for happy path vibecoding given small blast radius and LLMs can do that sort of static page in front of python backend really well.
I've also been surprised how much active learning I'm doing despite specifically not look at code. Between the need to spec things out carefully (plan.md) and fast iteration loop it's been a huge boost. Having the LLM look at a plan.md and suggest improvements has lead to a lot of "oh I didn't think about that" learning on architecture and user requirements link.
Presumably much of that learning boost is because I'm a hobbyist tier programmer, guessing professionals wouldn't experience the same since they learned this via manual coding trial & error over years.
I’ve been making stuff like this for a while (pre-LLM days). They are great little side projects that tend to be useful without getting overly complex. While I’ve vibe coded a couple just to get something done that I needed, I tend to like to write them myself still, as I enjoy the process.
> The alternative to CDNs is to use npm and have a build step for your projects. I find this reduces my productivity at hacking on individual tools and makes it harder to self-host them.
No. You can vendor these scripts & host them 1st party so you aren’t leaking data to these CDNs or risk users not actually getting the scripts. It isn’t like CDNs give you a performance boost anymore.
HTML tools are a good name. I called them something like "a local html file"
One problem I solved with this was a packer needed to scan a few (10-40) ids into his barcode scanner. It was not enough where pulling up their bulk-id-uploader program but also too tedious to go to some "number to barcode" website.
Turns out, barcodes can be made from a google font!
You can just display a number using that font. Then hooked up a for-loop that's progressed by pressing the space bar: paste in IDs, scan first, space, scan next, repeat.
Unrelated, but I bought a walking pad, and because I often use it at home while working on the laptop, and I wanted a nice UI to track the speed/time etc, I just asked claude to do one:
Another useful pattern for certain types of app is to include a function that saves the HTML file to your local drive/memory as a new file - for example, if the app features user inputs like writing or drawing.
> "If you want to see the code and prompts, almost all of the examples in this post include a link in their footer to “view source” on GitHub. The GitHub commits usually contain either the prompt itself or a link to the transcript used to create the tool."
As if your steady stream of learning-in-public experiments and insights weren't generous enough. Seriously, massive kudos for sharing all the details.
Why not introduce a single shared CSS for style consistency? Not full CSS separation, each tool could still have its local CSS.
Things like styling buttons, responsiveness, and so on are better solved once.
A good rule of thumb is: if the shared CSS fails to load, page still fully works but it might be uglier (weird fonts, etc). That's a reasonable rule for proper isolation (tools remain simple to understand, code remains reusable, etc).
I love the idea of self-contained tools, but you're already using CDNs. Having a shared CSS wouldn't hurt and actually make the tools better.
I would go as far as having a shared JS too (same idea, works if it doesn't load).
Each spiral is mostly independent. You can go ahead and delete the shared CSS from the <head>, they still work and don't break funcionality. However, by having the shared CSS I made them consistent, made them friendly to phone users and so on.
One tool I'd really like to see in this format is a simple "turn the background of this PNG to transparent". Models still refuse to follow the instruction to create transparent backgrounds for logos they create, and I often have to look for other tools doing this as post-processing.
It's possible that this is too complicated for the "few hundred lines of js" code envelope, though.
In the same spirit I have started building single page nanodjango +htmx apps hosted on railway, starting with, you guessed it, a Todo list app. No login needed.
Surprised AI studio isn’t recommended here - you can go from prompt->preview->deploy in like a couple seconds.
They have a library of sample apps you can edit but I wish they included the prompts and history to build each since I generally can’t get large apps to work - after a while the I’ll just produces more bugs as complexity grows. But I’m also a bad vibe coder and never read the code so entirely my fault :)
58 comments
[ 5.8 ms ] story [ 76.2 ms ] threadThis really showcases the power of the single page apps and why web will be always ahead of native for this kind of Swiss Army Knife tools.
With LLMs, it gets ridiculously easy to “develop” (generate) those too.
More recently, I've found a lot of benefit from using the extended thinking mode in GPT-5 and -5.1. It tends to provide a fully functional and complete result from a zero-shot prompt. It's as close as I've gotten to pair programming with a (significantly) more experienced coder.
One functional example of that (with 30-50% of my own coding, reprompting and reviews) is my OntoGSN [1] research prototype. After a couple of weeks of work, it can handle different integration, reasoning and extension needs of people working in assurance, at least based on how I understood them. It's an example of a human-AI collab that I'm particularly proud of.
[1] Playground at w3id.org/OntoGSN/
I've also been using LLMs to create and maintain a "work assist" Chrome extension that I load unpacked from a local directory. Whenever I notice a minor pain point, I get the LLM to quickly implement a remedy. For example, I usually have several browser tabs open for Jira, and they all have the same company logo as the favicon, so my Chrome extension changes the favicon to be the issue type icon (e.g. Bug, Story, etc) when the page loads. It saves a little time when I'm looking for a specific ticket I've already opened.
Not sure why, but the moment the file is split into files and subfolders, coding agents tend to do a lot more changes that what is absolutely necessary. That way a single html file wins!
My tool collection [0] is inspired by yours, with a handful of differences. I'm only at 53 tools at the moment.
What I did differently:
Hosted on Cloudflare Pages. This gives you preview URLs for pull requests out the box. This might be possible with Github Pages but I haven't checked. I've used Vercel for similar projects in the past. Cloudflare seems to have the odd failed build that needs a kick from their dashboard.
Some tools can make use of Workers/Functions for backend processing and secrets. I try to keep these to a minimum but they're occasionally useful.
I have an AGENTS.md that's updated with a Github action to automatically pull in Claude-style Skills from the .skills directory. I blogged about this pattern and am still waiting for a standard to evolve [2].
I have a base stylesheet that I instruct agents to pull in. This gives a bit of consistency and also let's them use Tailwind, which they'd seem to love.
[0] https://tools.dave.engineer/
[1] https://github.com/dave1010/tools/tree/main/functions
[2] https://dave.engineer/blog/2025/11/skills-to-agents/
Edit: come to think of it, I should revisit it now that everyone can vibe code. The sheet was to allow people to add to it, now maybe easier for me to take a message and ask an agent to update the html directly
Couple of unsolicited comments: first is that on mobile, the featured badge sits on top of the right facing arrow. Second is that the bubble level seems to be upside down? The bubble sinks rather than floats at least on my pixel
Sorry if this sounds overly critical, but what do you mean "only at 53 tools?" Was there a memo I missed about a competition to host LLM-built tools?
[1] https://labs.google/disco
[2] https://dl.acm.org/doi/10.1145/3706598.3713285
Personal tools seem like a reasonable place for happy path vibecoding given small blast radius and LLMs can do that sort of static page in front of python backend really well.
I've also been surprised how much active learning I'm doing despite specifically not look at code. Between the need to spec things out carefully (plan.md) and fast iteration loop it's been a huge boost. Having the LLM look at a plan.md and suggest improvements has lead to a lot of "oh I didn't think about that" learning on architecture and user requirements link.
Presumably much of that learning boost is because I'm a hobbyist tier programmer, guessing professionals wouldn't experience the same since they learned this via manual coding trial & error over years.
No. You can vendor these scripts & host them 1st party so you aren’t leaking data to these CDNs or risk users not actually getting the scripts. It isn’t like CDNs give you a performance boost anymore.
https://httptoolkit.com/blog/public-cdn-risks/
https://github.com/blue-monads/potatoverse
I wonder if packaging the results as web components would be the next logical step.
One problem I solved with this was a packer needed to scan a few (10-40) ids into his barcode scanner. It was not enough where pulling up their bulk-id-uploader program but also too tedious to go to some "number to barcode" website.
Turns out, barcodes can be made from a google font!
https://fonts.google.com/specimen/Libre+Barcode+39
You can just display a number using that font. Then hooked up a for-loop that's progressed by pressing the space bar: paste in IDs, scan first, space, scan next, repeat.
https://pastebin.com/5HRLh1G6
it does something like this
https://imgur.com/a/888BtpG
and connects through BLE
I found out about a new Python HTML parsing library - https://github.com/EmilStenstrom/justhtml - and wanted to try it out but I'm out without my laptop. So I had Claude Code for web build me a playground interface for trying it out: https://tools.simonwillison.net/justhtml
It loads the Python library using Pyodide and lets you try it out with a simple HTML UI.
The prompts I used are in this PR: https://github.com/simonw/tools/pull/156
Create PDFs from images, a Wordle hint/solver, or a classic DVD screensaver. Lots of stuff.
As if your steady stream of learning-in-public experiments and insights weren't generous enough. Seriously, massive kudos for sharing all the details.
Things like styling buttons, responsiveness, and so on are better solved once.
A good rule of thumb is: if the shared CSS fails to load, page still fully works but it might be uglier (weird fonts, etc). That's a reasonable rule for proper isolation (tools remain simple to understand, code remains reusable, etc).
I love the idea of self-contained tools, but you're already using CDNs. Having a shared CSS wouldn't hurt and actually make the tools better.
I would go as far as having a shared JS too (same idea, works if it doesn't load).
That's essentially what I did in https://alganet.github.io/spiral/ (also vibe coded).
Each spiral is mostly independent. You can go ahead and delete the shared CSS from the <head>, they still work and don't break funcionality. However, by having the shared CSS I made them consistent, made them friendly to phone users and so on.
One tool I'd really like to see in this format is a simple "turn the background of this PNG to transparent". Models still refuse to follow the instruction to create transparent backgrounds for logos they create, and I often have to look for other tools doing this as post-processing.
It's possible that this is too complicated for the "few hundred lines of js" code envelope, though.
https://web-production-1fc69.up.railway.app/
https://www.hackyexperiments.com/micro
They have a library of sample apps you can edit but I wish they included the prompts and history to build each since I generally can’t get large apps to work - after a while the I’ll just produces more bugs as complexity grows. But I’m also a bad vibe coder and never read the code so entirely my fault :)