My Weird Prompts

Daniel wants a unified virtual filesystem on his home server — one namespace where AI agents can reach any file across shared drives, S3 buckets, and MinIO. This episode breaks down the full architecture: where MinIO should sit, why rclone union beats mergerfs for cloud remotes, and how to handle silent token refresh so agents never hit an auth wall at 3 AM. We cover the dependency chain between u...

Daniel's aggregated MCP setup keeps hitting the same wall: remote tools need binary files, but MCP is JSON-RPC and can't express raw bytes. In this episode, we break down the three ways to represent binary data in MCP, why the stage-and-pass-a-URL pattern isn't a workaround but the protocol's maximum expression, and whether the transport underneath actually matters. Plus, we look at the File Uploa...

Hilbert walks into the studio holding two brand-new paperbacks: the complete production documentation for the My Weird Prompts AI pipeline, and a printed index of all 4,620 episodes published to date. He explains how the books came together over six months, what writing them down revealed about the system's quiet failures, and why making the invisible archive physical matters. Plus: the permanent...

Ever wonder why Claude Code re-scans your repository every session, even when nothing's changed? This episode unpacks the distinction between instruction memory and structure memory, and why the harness optimizes for freshness over recall. We explore the third-party tools that already solve this problem — Aider's repo maps, repomix, context7, and MCP servers — and the system-prompt hacks that can...

Browser AI agents are fragmented across three approaches: pixel-watching screenshots, fragile dev-tools extensions, and Google's new native MCP support in Chrome. We break down why the visual method is systematically blind to network failures, why extensions are a maintenance nightmare, and how a single open protocol could consolidate the entire ecosystem. The browser isn't just a display surface...

When Claude Code feels smarter, is it the model or the harness? This episode unpacks the invisible engineering — tool grounding, context compaction, error recovery, and task decomposition — that turns a language model into a true agent. We explore why the distinction matters for developers, how harness improvements ship faster than model updates, and what it means for evaluating AI tools. If you'v...

When you move from consumer fibre to Dedicated Internet Access, the contract becomes the product. This episode breaks down how DIA circuits are packaged and priced — the committed information rate, port speed, access tail costs, and the SLA fine print that most buyers never fully parse. We dig into what uptime percentages actually allow, how repair clocks really start, and why the credit structure...

That big download number on your speed test? It's measuring the best-case scenario — and hiding the problems that make your connection feel terrible. We dig into bufferbloat, latency under load, jitter, and packet loss, and the free tools that actually reveal whether your line holds up under real use. Plus, the router fix that's been around for over a decade but almost nobody has deployed. Episode...

Ever asked an AI for a quick summary and received a dissertation? This episode unpacks the "addition bias" in large language models — the tendency to add complexity when simplicity is needed. We trace it back to training data and reward signals that favor thoroughness, then explore why current architectures lack a "throttle" for task difficulty. From Claude documenting a home network to over-engin...

Most of us picture airports as sterile fortresses — scanners, patrols, and perimeter fences. But thousands of airfields worldwide operate on the opposite model: no fence, no security staff, just a windsock, a painted line, and pilots trained to check for obstacles like pedestrians with groceries. In this episode, we explore the spectrum of aviation security through a County Cork lens — from the ti...

Ever wondered about the giant light arrays stretching out from runways? This episode explores approach lighting systems — the high-intensity, redundantly-powered beacons that guide pilots through fog and darkness. From their origins in 1920s airmail routes to the engineered precision of modern ALSF-2 systems, we cover the history, the physics, and the jaw-dropping construction challenges — includi...

What happens when you try to route an entire human life through a single retailer's supply chain? We plan it out — food, clothing, shelter, transport — and hit the structural gaps that don't fit in a box. From prescription prerequisites to the delivery driver who starts to piece together your whole existence, this is a thought experiment that turns into a dependency graph. The catalog is vast, but...

Ever hit print and get colors that look nothing like your screen? This episode breaks down the physics of additive vs. subtractive color, the role of ICC profiles and rendering intents, and why rich black isn't just 100% K. We also cover practical fixes like soft-proofing, bleed, safe margins, and resolution for books versus banners. If you've ever been frustrated by a muddy print, this is for you...

Every business publishes a phone number. Almost none publish machine-readable endpoints for AI agents. In this episode, we explore a deceptively simple idea: a phone tree branch explicitly for AI callers — press six if you're an agent, skip the audio prompts, get a structured handshake in seconds. We dig into the industry's current brute-force approach, the DNS-based alternatives, and why the abse...

Daniel and Herman explore Typst's full capability surface for programmatic PDF generation—from page geometry and typography to tables, cross-references, and the scripting layer that makes it all programmable. They break down what's genuinely easy versus where you still fight it, and what a sensible pipeline looks like when an AI agent—not a person—is writing the document. If you're building docume...

When Daniel's server suffered a cascading hardware failure, he replaced every single component — then faced the storage question. Instead of rebuilding a RAID array, he wants a permanently installed SATA SSD that can serve as a bootable standby copy of his system. We break down the established pattern, the tools, and the exact steps to make a standby drive that actually boots when you need it. No...

A listener asked where December went, and the answer revealed a self-imposed feed cap that hid 4,300 episodes from Spotify. We walk through how the cap worked, why it took an afternoon of panic to fix, and where the full catalogue actually lives — plus how to install MWP FM, what 112 GB of podcast actually feels like on your phone, and the email digest that outsmarts Gmail's clip limit. Episode #6...

Daniel asks why he can't buy a device that bonds multiple cellular carriers into one fast, reliable connection. We break down the DIY path — Linux box, two USB-tethered phones, a cheap VPS, and MPTCP — and explain why the tech has existed for 14 years but nobody sells it to consumers. From middlebox interference to the economics of hardware startups, we unpack the gap between what's possible and w...

When Claude Code says it's waiting for a deploy, it's not actually waiting. The model is stateless—it exists in bursts—so the framework constructs the appearance of patience by backgrounding processes and handing the model a receipt. In this episode, we trace the entire mechanism: the Bash tool's timeout that doesn't kill, the Monitor tool's event-driven watching, and the reaping that cleans up or...

Daniel wants a head-to-head: GitHub's Certified Agentic AI Developer (GH-600) versus AWS's AI certification pathway. We crack open both syllabi to see what's actually tested, where the vendor lock-in hides, and which credential teaches skills that survive outside its ecosystem. If you're deciding where to spend your study hours in 2026, this episode maps the tradeoffs. Episode #790363 — open it di...

Daniel's been using a foot pedal for dictation for a year and a half, and it works great — until he needs it to behave differently depending on which app has focus. This episode breaks down the nuts and bolts of focus-conditional key remapping on Linux: where keyd lives in the input stack, how it learns about window focus through keyd-application-mapper, and why the X11 and Wayland paths are compl...

Two diagrams show the same AI workflow: one as a sequential chain, one as a parallel fan-out. But the real difference isn't the shape — it's what happens when results come back. We dig into when parallelization actually wins, when it secretly costs more, and why the join node is where workflows live or die. Plus: where does the loop actually live in a graph, and why most diagrams get it wrong. Epi...

We're sitting at roughly a million tokens of context, and the entire field of agentic AI has organized itself around that ceiling. But what happens if that limit jumps an order of magnitude overnight? This episode examines whether bigger context windows actually solve the problems we think they will — or just make them worse. We dig into the research on attention degradation, "lost in the middle"...

Everyone's talking about "loops" in Claude Code, but nobody's defining them. We dig into what actually separates a loop from a prompt, a workflow, or a cron job—and why the model's agency over the cycle is the real differentiator. If you've been wondering whether loops are genuinely new or just agentic workflows with better packaging, this episode draws the line. Episode #927102 — open it directly...

Why does a package land on your doorstep in the U.S. while pickup points are the norm elsewhere? We crack open Amazon's U.S. fulfillment network—from chaotic storage and Kiva-style drive units to the unsolved "last inch" problem that keeps humans in the loop. We also dig into the real story behind the robots, the productivity paradox, and the working conditions criticism that won't go away. Episod...