SaaS That App - Building Tech-Enabled Businesses

There are genuine practitioner nuggets buried in the episode—Claude skills as repeatable objects, markdown file security attack vectors, bounding-box confidence flagging for document review, Laravel ecosystem plugins improving Claude's domain awareness—but they are swimming in extended Scout project descriptions, a prolonged nature/Uber Eats tangent, and merit badge trivia that add nothing for a B2B operator.

“in your claw MD file, you just say, Everything must have a test and put it in all caps. And all of a sudden it just happens. I think I just passed 1,000 tests in this app.”

“I want you to give a confidence rating to everything that you are bringing in. And anything that is dipping beneath 80% confidence, I want you to flag it”

The observation that CLAUDE.md markdown files are a wide-open attack surface for prompt injection is a genuinely fresh security framing, and the 'object permanence for a thread' metaphor for skills is a useful conceptual reframe; but the bulk of the episode—ChatGPT-to-Claude migration arc, hallucination complaints, plugin marketplace quality decay—is a story that has been told many times in many podcasts.

“prompt injection is a feature, not a bug. You will never fix it”

“the markdown file of a skill markdown file or claude markdown file is now just like an open door to just completely compromise a system”

Steve Powell is a genuine builder who shipped a live, user-facing application and can speak concretely to architectural decisions, but he is a small-company owner and friend of the hosts experimenting with AI tooling—not a senior operator who has scaled something significant, and the insider dynamic limits the critical distance needed for high-caliber insight extraction.

“I had a functional system built in under 200 hours. And for me that's crazy.”

“The system is live. It's up@myeaglecoach.com and I actually have actual Scouts using it right now”

The episode punches above average on concrete data points—200 hours to full build, 100 hours to basic function, 12 seconds versus 25 minutes for document review, 1,000 tests at 100% coverage, Lighthouse score from 85 to 99, an 80% confidence threshold for human-review flagging—giving listeners real benchmarks to compare against their own efforts.

“I had a functional system built in under 200 hours...At about 100 hours it was ready to start going pretty basic”

“Something that took 20, 25 minutes, took 12 seconds.”

The hosts do redirect tangents and ask reasonable follow-ups (bounding box workflow, the skills-versus-thread distinction), but the overall dynamic is a friendly BS session between colleagues—no hard questions about monetization, security architecture specifics, or where the approach breaks down, and several tangents (nature walk, Cub Scout hand signs) are allowed to run unchallenged and unproductively.

“talk about that just a second, Steve, because I think we're approaching a point where people are starting to understand that there is some segmentation”

“Are you doing any bounding box work with that bounding box?”