For my stateful agent, I run a private, real-time dashboard on top of a Cloudflare durable object. This allows me to monitor tool calls and progress without needing the agent to constantly send messages to my chat client. These events are read and published directly from logs.

My stateful agent has been deploying websites and apps it builds during its autonomous work blocks. Before this, it posted interactive pages to its blog. Before that, it wrote posts and before that, scripts and simulations on its VM. Usually, I need to show it the way before an architectural leap.

I’m experimenting with building, persisting, and running code itself from ATProto. Most apps use ATProto to store content for the app. This approach treats the code as content. The app can be loaded using a viewer that anyone can implement, login with Bluesky, storing data as ATProto records.

Experimenting with a stateful agent has opened the door for a new form of software development that occurs entirely from my phone. The agent has dedicated GitHub and Cloudflare accounts and deploys cross-platform, personal apps that store data to my iCloud www.danielcorin.com/posts/2026/p...

Maybe it's just me but I really enjoy making chat components to narrate language model conversation for my blog posts and in general. This has been one of the deepest forays into design that I've taken, trying to get these to look just how I had in mind.

Throwback to typing out every character of my code. It was a simpler time ampcode.com/how-to-build...

I’ve been playing with an ambient chat concept, so I designed and generated a bunch of different chat UIs to see what felt right www.thoughteddies.com/notes/2025/l...

Having done a lot of vibe coding lately, I think I’ll move away from it exclusively. Current agents eventually fail to adhere to some prompt despite various attempts and approaches. Whenever this happens and I look in the codebase I am usually mortified by what I find.

I write convenience scripts to remove cruft in dev workflows. These start with one off scripts that I parameterize and document. With tools like Goose, LLM use cases no longer require you to write the script. Just write the spec and let the model execute www.danielcorin.com/posts/2025/g...

A note taker that generates various different "indexes" (summaries and similar) using an LLM and different target templates

What happen to completion models like `text-davinci-003`? Why does no one train or release these types of models anymore? Maybe I am just romanticizing early models, but I recall this model being much more steerable that chat models because there was no bias to responding conversationally

Can someone point me to an inpainting API they like? I am using Flux Canny served by together.ai. Ideally a low-latency Flux Inpainting API or something similar.

Hoping to be “wrong on the internet” and have someone tell me how to workaround the triple backtick stop sequence in Cursor www.danielcorin.com/til/cursor/t...

Vibe coding is tantalizing in that it often can feel like you’re one prompt away from solving a problem when you are not. I’ve wasted plenty of time rolling the dice with the Cursor agent when I could have just dug in and done the work (even LLM assisted in a lower level way) and finished faster.

I’ve been experimenting with feeding around 200 log posts from my site into an LLM to summary, following a certain template structure. The result have been pretty promising with Haiku and Sonnet. With a bit of prompt eng, Gemini also does well.

I open sourced my app, Delta, a UX for local-first conversation branching with LLMs. It let's you "rewind" and take conversations in different directions. You can also swap models, temperature and top_p to see how the model responds changes you make and support images. github.com/danielcorin/...

I hadn't yet had the time or need to prompt an LLM to cite its answers from the documents in its context but today Anthropic release will do it for you. Here's Python code to get you off the ground and running since I couldn't find something to quickly copy+run www.danielcorin.com/til/claude/c...

Currently playing around with a prompt+sketch as an input to an image-to-image model. It feels like cheating but it's incredibly effective for translating initial ideas out of your head.

Playing around with D3 and conversation branching. Here is a a conversation with several separate branches

Playing around with D3 in React/Astro to create from UX experiments for conversation branching. This attempt didn't quite go as planned.

Tailwind/React are a potent combo when building UIs with LLMs. Why is that? Sonnet seems biased towards this stack. I’d guess this is due to training. However, it also feels like co-location of style and structure would make it easier to effectively steer the model’s output through prompting.

Shamelessly consolidating some replies from this thread to showcase some apps/prototypes with LLM conversation branching UX

Who is building an idea and latent space explorer? Sort of a hybrid of a thought partnership and knowledge web traversal. I’m still mulling on what the UX could look like but I think it involves conversation branching, semantic zoom and some kind of topic clustering.

I’ve found software engineers to be surprising holdouts in adding language models to their workflows. I’m unsure if this is due to some kind of preference for probable correctness that comes with the territory of writing code or something else. Often I hear “the model gets it wrong (isn’t perfect)”

Can someone recommend me a simple, no-cloud-required Typescript LLM evals framework or should I just write what I need myself?

I’m not convinced it will be possible to rely on different models/providers for redundancy for a product use case. We can write evals and run across multiple models to validate parity but as model APIs seem to be diverging, I wonder if we’ll be able to avoid committing to a single provider.

Great writeup from Simon as always. I recently used LLM-generated code to debug a similar AWS networking issue: testing TCP, ping and telnet connectivity, SSL certificate verification, and even constructing a full request body from scratch in binary with just a short prompt to the model.

A day in the mind of Claude Sonnet

These are pretty nice My pelican on a bicycle prompt is intended to be deliberately absurd and hence very difficult to produce anything remotely "good", so it's fun to see that SVG sunsets can produce quite pleasing images

Inspired by @simonwillison.net's pelican-bicycle repo, here are some CSS sunsets by gpt-4o, claude-3-5-sonnet and gemini-2.0-flash-exp. Sonnet and Gemini animate the birds and clouds. Prompt: Generate pure HTML/CSS art of an extremely detailed, beautiful sunset. No talk or code fences. Code only.

(Finally) managed to get an MVP of conversation branching working. I have a bunch of ideas for where I want to take this project, but for the moment, I'm quite happy it's in a minimally shareable state. If you end up trying it out, let me know how it goes gist.github.com/danielcorin/...

Anyone have the secret codes to package better-sqlite3 into an Electron application built with vite?

I like Vercel `ai` SDK API a lot for streaming responses from language models to a frontend. Currently trying to replicate this API in Electron IPC. Coming from mostly React on the JS side, Electron is a whole new world to learn sdk.vercel.ai

Generated words from a several categories with a language model, then created embeddings for the words and plotted them to explore how they cluster

Another experiment -- early and rough. Generating a mindmap from a conversation as it progresses. It might make more sense to start with a concept and build a map that is clickable to drill further into a topic. Building the map incrementally is better since generation latency is an issue

I wrote about LLM conversation branching and how you can refine your ideas using a model as a sounding board www.danielcorin.com/posts/2024/c...

Multiple PhDs who I’ve encouraged to try frontier models for their software needs have reported positive results with excitement to try more things. For most folks, convincing them to try it out is the most difficult part. The utility becomes undeniable when you see it for yourself

An experiment with a story builder type UI. Content by Sonnet, image by Dalle 2. This one has been harder to iterate on compared to the conversation branching UI for some reason. I struggled to get the model to restrict the viewport overflow-y. I'm probably done with this one for a while

Claude/Cursor let me move fast compared to what I’m used to when building frontends. There are definitely times I get stuck, but the combination of code gen, chatting and copy/paste gets me far. However, my learning of the details has plateaued. I still have learned very little Tailwind.

Some screenshots I took throughout development. I'm planning to do a better job of documenting progress like this in the future. It's gratifying to see where things started compared to where they ended up. bsky.app/profile/danc...

Playing around with LLM conversation branching for idea exploration and learning