Sebastian Schkudlara
The Danger of "Sloppypasta"
Lately, it feels like I’m drowning in “sloppypasta.” It’s incredibly easy to ask a chatty AI assistant living in your IDE to solve a problem, and then blindly copy-paste thirty lines of fragile boilerplate directly into a production codebase. It’s so easy to trade architectural rigor for the convenience of instant code generation.
While IDE integrations are undeniably useful, treating an LLM purely as a conversational co-pilot can sometimes lead to bloated, loosely generated code. I’ve found that there’s a different, highly effective way to work with these models that enforces discipline: pushing the interaction down to the terminal level.
The Unix-Agent Paradigm
I honestly believe we should treat LLMs exactly as they were meant to be treated: as composable UNIX pipes.
By forcing AI outputs through standard CLI streaming pipelines, we strip away the conversational filler and demand raw, structured data. When you bind an LLM to standard I/O and pipe it into existing tools, you enforce strict structural constraints. The model stops being an unpredictable AI and becomes a tightly-scoped, easily auditable background process.
Building for Terminal-Level Constraints
I believe in this approach so much that I’ve been heavily investing in it. I recently spent cycles optimizing the CLI streaming pipelines in switchAILocal for lower latency and hardened the subprocess I/O handling just to support these workflows.
Furthermore, my recent integration of the Cortex Natural Language Interface (Phase 3) into the traylinx-cli is built entirely on this foundation. By strictly managing raw subprocess I/O and enforcing Natural Language Interfaces directly at the CLI, the LLM turns into a sharp, single-purpose utility.
If you want to write software with LLMs without slowly corrupting your codebase with unvetted boilerplate, try piping the model through your terminal. You’ll enforce standard constraints, get cleaner outputs, and get back to actual engineering.