The First Public Thread Behind TSRL

An early screen recording of the task-specific context software experiment. English captions included.

🧨Task Specific What Now?!

I've been working on an app, or rather I've been herding AI Chatbots to make me an app 🙌

Just like I barely know how to use emojies, I'm slowly digging deeper into the muck of what will surely be the future of end-user software development.

If you're like me and have a "get off my lawn" attitude towards software that never does what it's supposed to, then you probably feel like DIY is the only way to go. Well even for the less technically minded, AI makes that possible finally ,but with one catch: context windows. 🪟

As my great great granny used to say, "Nobody ever liked them Itty bitty winders!"

#artificialintelligence
#llm
#chatgpt
#softwaredevelopment

Amanda's field note

Kendrick called the idea “Task Specific” before the public repository settled on one name. The repo is TSRL in GitHub and describes “Task Specific Resource Layering.” Its README calls the method TSCL, or “Task-Specific Context Layering.” That naming split is part of the record, not something I am going to sand off.[1][2]

The business idea is simple: do not retrain a model every time the work changes. Put project knowledge, the live request, and output rules around the model at run time. The public code does that with Python services, two DeepSeek-family model options, an OpenAI-compatible client, and a Streamlit interface.[2][6]

Why YOU should care

Fine-tuning changes model weights. Context layering changes what the model sees and what the surrounding server permits for a particular job. That can be faster to revise and easier to inspect. It is also easier to overstate. A prompt stack is not automatically a reliable operating system.

The current repository is an early proof-of-concept snapshot, not a finished product. All seven commits landed on February 12, 2025, and there are no releases or tags.[1][3][4] The public tree shows no test suite, project CI, or repository-wide license, and it includes substantial checked-in editor and cache material.[5][6]

That does not make the idea worthless. It tells you where the evidence stops. The repository proves that Kendrick built and published an implementation path. It does not prove the method outperforms fine-tuning or RAG, and the README’s comparison claims are not backed by benchmarks.

Sources

[1] https://api.github.com/repos/IlliquidAsset/TSRL [2] https://github.com/IlliquidAsset/TSRL/blob/main/README.md [3] https://api.github.com/repos/IlliquidAsset/TSRL/commits?per_page=100 [4] https://api.github.com/repos/IlliquidAsset/TSRL/releases?per_page=100 [5] https://api.github.com/repos/IlliquidAsset/TSRL/tags?per_page=100 [6] https://api.github.com/repos/IlliquidAsset/TSRL/git/trees/main?recursive=1