The command line is no longer just for developers. Researchers, writers, and analysts are turning terminal-based AI agents into powerful productivity workspaces—without writing a single line of code. From managing equity research to organizing personal therapy notes, these "non-coders" are redefining what these tools can do. We explore the three pillars making this possible: repo-as-workspace, persistent instructions, and MCP servers that connect to the real world.

Episode #353570 — open it directly at myweirdprompts.com/353570

The podcast explores the architectural tension between visual "Computer Use" agents—like Anthropic's demo—and API-first automation. Hosts analyze whether visual agents are a high-latency bridge to a protocol-driven world or a necessary tool for legacy systems. They discuss cost implications, reliability issues, and the potential for visual interaction to become just another capability rather than a standalone product category.

Episode #542828 — open it directly at myweirdprompts.com/542828

Every day, companies lose massive amounts of institutional intelligence because AI chat outputs are treated as disposable. In this episode, we explore the "ephemeral context trap" — the gap between brilliant AI conversations and permanent knowledge bases. We discuss why current tools fail to capture the "trail of thought," and outline a five-step pipeline (Capture, Sanitize, Extract, Categorize, Human-in-the-Loop) to turn ephemeral chats into structured, searchable assets. Plus, a look at tools like Dust, Khoj, and Microsoft Presidio that are building the plumbing between generation and storage.

Episode #748252 — open it directly at myweirdprompts.com/748252

What happens to your AI's brilliant answers after you see them? In this episode, we explore the "leaky bucket" problem of AI output storage. We discuss why treating AI conversations as ephemeral is a corporate nightmare, and dive into the tools trying to give these models a long-term memory. From LangSmith and Langfuse to "Reverse RAG" and projects like Mem zero and Letta, we uncover how to turn a mountain of raw logs into a goldmine for fine-tuning and compliance. We also examine how temporal awareness and automated evaluation are creating smarter, more stateful AI partners.

Episode #168110 — open it directly at myweirdprompts.com/168110

We move past vague ethics to the literal plumbing of AI safety. This episode explores the specific libraries, proxy layers, and architectural decisions that act as the new enterprise firewall for LLMs. We dissect the tension between latency and security, comparing "sandwich" guardrails with token-level steering, and break down the open-source versus commercial landscapes—from NVIDIA NeMo and Guardrails AI to Lakera's threat intelligence.

Episode #137032 — open it directly at myweirdprompts.com/137032

We have massive AI models that claim to be "world-class intelligent," yet they often fail at basic tasks like finding a specific fact in a long document. This episode explores the disconnect between benchmark scores and real-world performance, diving into EvalScope, an open-source toolkit designed to stress-test long-context retrieval and agentic capabilities. We discuss the "lost in the middle" phenomenon, the danger of overfitting to public benchmarks, and why testing speed and tool-use is just as important as raw intelligence.

Episode #712749 — open it directly at myweirdprompts.com/712749

The industry is scaling faster than humans can review, so we’ve turned to LLM-as-a-Judge to grade model outputs. But this creates a hall of mirrors: AI grading AI, often with a preference for verbosity and its own style. We explore the mechanics of single-point, pairwise, and reference-based scoring, and the hidden biases—like position and self-enhancement—that threaten to create a monoculture of identical models. Is this the future of evaluation, or a trap we can’t escape?

Episode #508807 — open it directly at myweirdprompts.com/508807

We all know how to test if an LLM solves a math problem, but how do you measure if it has the right "soul"? This episode tackles the messy world of qualitative AI evaluation. We explore why binary benchmarks fail for real-world tasks like medical summaries or creative writing, and dive into techniques like LLM-as-a-Judge, G-Eval, and counterfactual testing to map a model's hidden worldview. Learn how to build a "Golden Dataset" and avoid the pitfalls of subjective bias.

Episode #405001 — open it directly at myweirdprompts.com/405001

How do you know if your LLM is actually working? This episode moves past gut checks to explore the real metrics—coherence, hallucination, instruction-following—and why even your hardware choice can skew results.

Episode #786706 — open it directly at myweirdprompts.com/786706

As AI agents move from prototypes to production, teams face a fragmented mess of inference gateways, MCP servers, and observability tools that don’t talk to each other. This episode explores the rise of the "AI Control Plane"—a unified infrastructure layer that promises a single pane of glass for routing models, managing tools, and tracking costs. We dig into how these systems handle security, context, and tool namespacing, and why the industry is coalescing around terms like "Single-Origin AI Infrastructure." Whether you’re battling duct-taped scripts or planning an enterprise rollout, this is your guide to the plumbing that makes AI agents actually work.

Episode #191526 — open it directly at myweirdprompts.com/191526