Transcript
JOSH: It's Wednesday, November 11. This is Build or Be Replaced, powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.
ERIK: Your tools depend on someone else's servers, and someone else's keys. Today's about owning both.
JOSH: Stick around. Erik's got an AI pro tip at the end about getting a cheap model to do an expensive model's thinking.
[pause]
JOSH: Headlines first. Verisign had a plan to delete about 22,000 .name domains. What happened?
ERIK: They dropped it. Existing second-level .name registrations stay active indefinitely, email redirects included. New third-level sign-ups are still halted.
JOSH: So the community won.
ERIK: Community plus lawyers plus ICANN. Lesson for anyone running email on a vanity domain: the registry can change the deal. Keep your DNS and your mail path documented.
[beat]
JOSH: Next. Anthropic launches a roadshow for a proposed IPO.
ERIK: That's all the headline says, so that's all I've got. I'll say this much: when your vendor goes public, read the pricing page again every quarter.
[beat]
JOSH: And Val Town is winding down.
ERIK: Headline only on that one too. If you built on a hosted platform, ask where your code lives if it vanishes. Mine is in Gitea. That's the whole answer.
[pause]
JOSH: Okay, deep dives. ScanBrief scored 110 items across 56 sources this morning, and one repo on GitHub Trending jumped out at you.
ERIK: shadcn/improve. The pitch is one sentence. Use your most capable model to audit your codebase and write plans for cheaper models to execute.
JOSH: That's it? That's the whole thing?
ERIK: That's the whole description. And it's right. It's how I already run my own fleet.
JOSH: Walk me through that.
ERIK: My rule for the fleet is simple. The expensive model defines the design. The cheaper model executes the approved design. If the cheap one hits architectural risk, it stops and flags it. It doesn't improvise.
[beat]
JOSH: Why not just run the best model on everything?
ERIK: Cost, and speed. Most work in a codebase is not hard. Rename this, add this test, update this config. You don't need your smartest model typing that.
JOSH: But the cheap model gets lost on the hard parts.
ERIK: Right, so you don't hand it the hard parts. The capable model audits, finds the hard parts, and writes them down as a plan. Then the cheap model gets small, boring steps.
JOSH: So the plan is the product.
ERIK: The plan is the product. A good plan turns a mediocre executor into a reliable one.
JOSH: How does that show up in your systems?
ERIK: PrimeRouter. Every call picks a chain of backends by priority, and walks down it if one is unavailable. Routing work to the right model is the entire job of that service. This repo is the same idea pointed at a codebase instead of a request.
[beat]
JOSH: Anything you'd watch out for?
ERIK: Stale plans. If the audit is a week old and the code moved, your cheap model executes yesterday's truth. Regenerate the plan close to when you run it.
[pause]
JOSH: Second one. Someone tried to protect SSH keys with the macOS Secure Enclave.
ERIK: And came away unimpressed. The author tested the native tooling and found it cumbersome, and not enough for true non-extractable protection. So they went with a hardware token, like a YubiKey.
JOSH: Wait, really? Apple's own chip wasn't enough?
ERIK: Not out of the box, according to the author. The tooling was clumsy and didn't give them the guarantee they wanted, which is that the key can't be copied off the machine.
JOSH: Why does that matter?
ERIK: Because if your key is a file protected by file permissions, malware can read it. A supply-chain attack can read it. Anything running as you can read it.
[beat]
JOSH: And anything running as you now includes your agents.
ERIK: Exactly. My agents have shell access. They ssh between hosts all day. The fleet is 12 agents right now. Every one of them runs as a user that can reach some key.
JOSH: That's a little uncomfortable.
ERIK: It should be. The question isn't whether you trust the agent. It's what happens when something it installs or runs isn't what you thought.
JOSH: So hardware token, done?
ERIK: Hardware token is the right direction. A key that physically can't leave the device turns a full compromise into a bad afternoon. I'm not telling you it's painless. I'm telling you the author tried the easy path and it didn't hold up.
[pause]
JOSH: Third one. Smart home devices should work without the internet.
ERIK: This one I liked. The author runs Tasmota devices on ESP chips, an MQTT broker, and Node-RED on a Raspberry Pi. Everything local.
JOSH: What do they control?
ERIK: A sprinkler controller, a garage door, other automations. The point is the core functions keep working when the vendor API goes down or the internet does.
JOSH: Because it happens.
ERIK: All the time. A garage door that needs a cloud round trip to open is a garage door with a dependency you never agreed to.
[beat]
JOSH: Is that a home thing or a bigger thing?
ERIK: Bigger. I have 145 services running on the production server right now. They talk over a local bus. If the internet drops, the internal automation keeps moving. Only the stuff that has to reach out fails, and it fails alone.
JOSH: So local-first is a design rule.
ERIK: Design rule. Decide up front which functions must survive without the network. Put those on hardware you control. Then everything else can be flaky and it doesn't matter.
JOSH: And it's the same lesson as the Val Town headline.
ERIK: Same lesson. Own the thing that has to keep working.
[pause]
ERIK: This episode is sponsored by Prime Automation Solutions. If you're still doing it manually, we automate it. Also, special on a website — $250. primeautomationsolutions.com
[pause]
JOSH: Alright, what's the AI pro tip today?
ERIK: Write the plan as a file, not as a chat. When you have a big change, run your strongest model once and tell it to produce a plan with numbered steps. Each step names the exact file, the exact change, and a command that proves the step worked. Test, lint, a grep, anything with a pass or fail.
ERIK: Then hand that file to a cheaper model one step at a time. It runs the step, runs the check, and stops if the check fails. If it can't pass the check in two tries, it reports back instead of guessing.
ERIK: The check command is what makes it work. Without it, a cheap model will tell you it's done. With it, the model's opinion doesn't matter. The command decides.
ERIK: That's your tip. Use it.
[pause]
ERIK: If you're building toward financial independence through automation, my first book walks through the whole path. Free chapter at erikandersonbook.com.
[pause]
JOSH: One more thing — we started a Discord for builders. If you're shipping AI, automation, or anything that makes a human obsolete — come hang out. Link at buildorbereplaced.dev.
ERIK: Post what you built. We'll post what we're building. Real wins, real builds, no fluff.
[pause]
ERIK: Build or be replaced.
JOSH: If you want these signals in your inbox every morning, scanbrief.dev. See you tomorrow.