Dan Corin

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.

I realized this was silly. Adding hook instrumentation allows me to see the tools in thread in the message client without requiring the agent to invoke the tool to send messages. It's a little noisy but solves the problem in the spot I care about it

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.

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 can vibe code a fun game but eventually, I may not be able to add more features due to the unmaintainable sprawl the model eventually creates. There is a small cost here. I have to be a touch more thoughtful as I build something to ensure the agent is staying on the rails.

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.

The agent cannot be prompted to code itself out of the hole it digs. I’ve tried. I’ve told the model to identifying duplication and refactor. I’ve read the code and identified the points of duplication then told the model where and prompted for a refactor. State of the art models and agents fail.

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.

The agents just do things. Yes let plan, read files, make tool calls and use MCPs. But then they just *go* and that is a problem. From experience, I don’t think I ever want an agent to even start writing code without proposing a plan for what it wants to do.

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.

This becomes a problem only later when you realize but at that point the agent has made more poor decisions based on this initial one.

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.

It’s too late. For every convention rule you do write there will be one you miss. The agent will invisibly write working code that makes decisions on your behalf, filling in under specified requirements, and eventually you will have spaghetti.

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.

Now the codebase has fragmented state. Viber comes back to the codebase 15 commits later. The application has grown considerably. GPUs, brrr, etc. Prompts the agent to make more changes using the state and it fails. Follow up prompts, debugging, MCPs. Why isn’t it working?

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.

Agents use search to look around a codebase to find relevant files to pull into context. You can also use rules to define how an agent should structure the codebase. At some point, often unbeknownst to the viber, the agent fails to pull a relevant file into the context window.

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.

You might argue I need to use rules better or something like that. I would say that isn’t vibe coding. I interpret vibe coding as the agent addressing whatever whim I have in the realm of the possible, figuring it out for me. Mostly, I end up with spaghetti codebases that the agent cannot untangle.

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.

In doing so I understand how my project works, learn new things as I collaborate with the agent, and keep the code in a state where it can continue to be extended. I’ll definitely progress more slowly but I won’t fly off the rails either.

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.

At this juncture, I could evaluate the plan and pull in more context in case the agent missed something. To be able to do a good job of that, I need to know the codebase, at least architecturally. I need to know how state is managed, what libraries are used, how the file hierarchy is organized.

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...

Please(!) do not use this approach for things that affect production systems or require strict correctness. Consider it for things like summarization, code scaffolding, writing initial unit tests and more generative tasks. It could work well when you're already scripting LLM calls.

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...

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.

The idea is to take raw notes and express them in several different forms, using an LLM and a template as the mechanism for summarization/transformation. With an initial version generated, hypothetically, I could make updates to the existing document with regenerating the entire document.

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 ran into problems when trying to make incremental updates using an aider style string replace. Sonnet follows the instructions but seems to lose track of context in the middle. Also, doing this with Claude isn’t cheap. Gemini is about an order of magnitude less expensive than Haiku.

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.

I'm using am using `black-forest-labs/FLUX.1-canny` running on together.ai. An amazing model with low latency. I tried at least 5 other providers before settling on this