My Weird Prompts

On September 11, 1974, Eastern Air Lines Flight 212 crashed 3.3 miles short of the runway in Charlotte. The cockpit voice recorder captured the crew discussing Watergate and used cars. The NTSB recommended a sterile cockpit rule six days later. It took the FAA seven years — and dozens more deaths — to write it down. This episode traces the rule from 14 CFR 121.542 through the crashes that kept re-proving it: Air Florida in 1982, Delta 1141 in 1988, USAir 1016 in 1994, Corporate Airlines in 2004, and Comair 5191 in 2006. Then it asks the uncomfortable question: if every one of those crews knew the rule, why did it keep getting broken? The answer lives in Robert Sumwalt's 1993 analysis of 63 ASRS violation reports, and in a brutal reframe — the rule isn't about whether crews talk. It's about when they stop.

Key sources:

• NASA ASRS Directline #4, The Sterile Cockpit, Robert L. Sumwalt, June 1993 — https://asrs.arc.nasa.gov/publications/directline/dl4_sterile.htm

• NTSB Aircraft Accident Report AAR-75-09 (Eastern 212), May 23, 1975 — https://www.faa.gov/sites/faa.gov/files/2022-11/Charleston%20EA212%20NTSB%20Accident%20Report%20AAR75-09.pdf

• Wikipedia, Sterile flight deck rule — https://en.wikipedia.org/wiki/Sterile_flight_deck_rule

Episode 5913 — myweirdprompts.com/n/5913 (permanent short link: myweirdprompts.com/789347)

Rita runs a two-person landscape design practice and wants a real backend instead of spreadsheets and rented CRMs. Can AI-assisted development make that realistic? This episode digs into the backbone: the database and architecture. It covers why the schema is the real moat, how to map a garden design practice onto the canonical CRM entities, and why a rigid schema with fast migrations beats a flexible one. Then it walks through the boring-but-correct stack — Cloud Run, Cloud SQL Postgres, Terraform — and the three things that actually break: connection pooling arithmetic, migration races at container start, and secrets management. Part one of a three-part series on building your own tools.

Key sources:

• Neon, Neon Pricing Plans , current as of 2026. — https://neon.com/pricing.md

• Google Cloud, Cloud Run pricing . — https://cloud.google.com/run/pricing

• Google Cloud, Cloud SQL pricing . — https://cloud.google.com/sql/pricing

Episode 5912 — myweirdprompts.com/n/5912 (permanent short link: myweirdprompts.com/605952)

A two-person architecture practice wants Revit 2027 running persistently so AI agents can query it through Autodesk's official MCP server. The hardware question looks simple until you price it out. We walk through Revit's actual system requirements, why single-core clock speed matters more than core count, what the read-only MCP tech preview can and can't do, and the three ways to arrange this setup — local workstation, modest box plus cloud rendering, or fully cloud-hosted. Along the way: Chaos's blunt warnings about vGPU and Enscape, why V-Ray's CPU rendering quietly changes the math, and where the budget actually lands.

Key sources:

• Autodesk — https://www.autodesk.com/blogs/aec/2026/06/17/revit-public-mcp-server/

• Autodesk — https://www.autodesk.com/support/technical/article/caas/sfdcarticles/sfdcarticles/System-requirements-for-Revit-2026-products.html

• Autodesk — https://www.autodesk.com/products/revit/buy

Episode 5911 — myweirdprompts.com/n/5911 (permanent short link: myweirdprompts.com/162633)

Michael runs a two-person architecture practice and wants Revit on an always-on VM so an AI agent can drive the model while both architects sleep. We break down whether that's actually buildable. Autodesk's official MCP server ships as three installers — 8 read tools, 11 write tools, 39 experimental — and not one of them runs arbitrary code. The community servers that do run arbitrary code are the ones with an unauthenticated TCP port open on every interface. Then there's the blind test: the native MCP counted 32 walls where there were 57, while a dumb IFC export answered every question correctly and four times faster. Plus the licensing question Autodesk hasn't answered, and why "always-on VM" doesn't actually give you a running Revit with a model open.

Episode 5910 — myweirdprompts.com/n/5910 (permanent short link: myweirdprompts.com/557315)

Rita runs a two-person business on Google Workspace and wants a real business layer on top of it. This episode walks through the integration architecture: why PostgreSQL owns the relationships while Google stays authoritative for documents, messages and events; how service accounts and domain-wide delegation actually differ; and why service accounts can't own Drive files — a constraint that forks the entire folder design. Plus the real costs of Cloud Run plus Cloud SQL, and what Google's new Workspace API quotas mean for small firms.

Episode 5909 — myweirdprompts.com/n/5909 (permanent short link: myweirdprompts.com/133931)

When a small landscaping practice bolts WhatsApp, a meeting assistant, and a CRM onto its existing backend, the instinct is to consolidate everything into one database. That instinct is wrong. This episode unpacks the distinction between a system of record and a source of truth, then walks through a concrete build order — WhatsApp Business Platform first, transcription second, CRM third — showing why every field needs exactly one declared owner and why copies are fine as long as you know who decides. Along the way: the new per-phone-number billing rules, the twenty-four-hour customer service window, why templates are copy that costs money and requires permission, and how to turn meeting transcripts into structured data instead of a folder of dead files.

Key sources:

• Meta, updated 2026-07-01 — https://developers.facebook.com/docs/whatsapp/pricing/updates-to-pricing/

• Meta, conversation-based pricing (deprecated) — https://developers.facebook.com/docs/whatsapp/pricing/

• Meta opt-in requirements — https://developers.facebook.com/docs/whatsapp/overview/getting-opt-in/

Episode 5908 — myweirdprompts.com/n/5908 (permanent short link: myweirdprompts.com/884398)

Daniel has a working desktop and an empty 4U slot in his rack, and wants to know what it actually takes to move one into the other. This episode works through the whole procedure: how rack units define your envelope (and why "4U" doesn't mean seven inches of usable room), why depth is the unstandardized dimension that sinks builds, and which three measurements — cooler height, GPU height and length, PSU depth — decide whether your parts come across. Then it's on to chassis selection: why workstation cases like Sliger, SilverStone, and Rosewill exist separately from server chassis, what to check before buying, and the real-world build logs where a Noctua NH-D15 turned out to be five millimeters too tall. If you've got a tower under the desk and an empty slot in the rack, this is the sizing conversation to have before you spend anything.

Key sources:

• Sliger , current product page. — https://www.sliger.com/products/cx4200a

• Sliger , current product page. — https://www.sliger.com/products/cx4170a

• RackChoice , 4U ATX chassis specs. — https://www.rackchoice.com/4U/37.html

Episode 5907 — myweirdprompts.com/n/5907 (permanent short link: myweirdprompts.com/281728)

What happens when you strip the national security justification out of compartmentalization and run an ordinary company on it? This episode takes a listener's Netflix-style pitch seriously — a workplace where nobody understands their own workload, colleagues can't discuss their work, and feedback comes from a manager whose identity you may never learn. The closest real precedent isn't a corporation at all: it's Oak Ridge, Tennessee, a city of 75,000 that didn't appear on a single official map for six years. We walk through what a typical day actually looked like there, why the company's own documents admitted the job couldn't produce satisfaction, and how the 1977 CIA "Critique of the Codeword Compartment" documented what happens when secrecy scales — including the rise of "collectors of clearances" who equated knowledge with power. The result is a structure that converts secrecy into unaccountable power, with or without a national security reason.

Key sources:

• CIA Center for the Study of Intelligence, Critique of the Codeword Compartment in the CIA, March 1977 (TR/IM 77-02J) — https://sgp.fas.org/othergov/codeword.html

• Conjecture, Information Hazard Policy, July 2022 — https://www.conjecture.dev/information-hazard-policy

• Alex Wellerstein, Oak Ridge Confidential, or Baseball for Bombs, *Restricted Data*, 2012-04-16 — https://blog.nuclearsecrecy.com/2012/04/16/oak-ridge-confidential-or-baseball-for-bombs/

Episode 5906 — myweirdprompts.com/n/5906 (permanent short link: myweirdprompts.com/590660)

OPNsense is a convenience, not a requirement — you can turn a cheap two-NIC mini-PC running Debian or Ubuntu into a full router with systemd-networkd, nftables, dnsmasq and WireGuard. Some people have done exactly that and never looked back: Ansible-provisioned configs, six hundred days of uptime, and patches straight from upstream instead of waiting on a handful of downstream maintainers. Others argue the sticker price of a router is only ten to twenty percent of its total cost of ownership, and the other eighty percent is the decade of maintenance nobody budgets for. This episode works through both sides — the patch-latency argument, the PPPoE performance case, the composability dream — and asks whether a "composable router" actually exists yet, or whether the closest thing to it is still in beta.

Key sources:

• MurOS repo, OPNsense fork on Debian 13, beta (README fetched 2026-10-12). — https://github.com/murosorg/muros

• (via fetch_repo) — https://github.com/petrouz/ansible-debian-firewall

• ezrouter (any systemd distro → router). — https://github.com/kalken/ezrouter

Episode 5905 — myweirdprompts.com/n/5905 (permanent short link: myweirdprompts.com/074234)

You've built the gateway group, set the tiers, and failover works. So why does a dropped link still kill your VPN, your game, and your VoIP call? This episode separates true bonding from load balancing, then walks through what the "advanced" version of multi-WAN actually looks like in open source. In OPNsense, that means gateway group tiers, trigger levels, weight-based distribution, sticky connections, and the one silent subnet-mask mistake that breaks per-device routing. On the OpenWrt side, it means mwan3's firewall-mark packet classification, the iptables-to-nftables transition, and a five-hundred-line Lua alternative called Mini-MWAN that does failover and nothing else. Along the way: why failover speed was never the bottleneck, why the public IP change is the real problem, and why people who've done load balancing properly still go looking at bonding.

Key sources:

• OPNsense docs, gateway groups/multi-WAN — https://docs.opnsense.org/manual/multiwan.html

• OPNsense docs, Multi WAN how-to (sticky, weight, failover states) — https://docs.opnsense.org/manual/how-tos/multiwan.html

• mwan3 nftables port, v3.6.12 (2026-08-11) — https://github.com/dl12345/mwan3

Episode 5904 — myweirdprompts.com/n/5904 (permanent short link: myweirdprompts.com/719827)