{"type":"rich","version":"1.0","provider_name":"Transistor","provider_url":"https://transistor.fm","author_name":"NC Tweener Talks","title":"[REDACTED] Episode 9: He Gave His AI Assistant 64 Tools and Zero Permission to Delete Anything","html":"<iframe width=\"100%\" height=\"180\" frameborder=\"no\" scrolling=\"no\" seamless src=\"https://share.transistor.fm/e/632fa0b4\"></iframe>","width":"100%","height":180,"duration":2565,"description":"Josh Guter, founder of Elephants, joins us for a live product teardown of the AI operations assistant he built for small teams. The conversation covers the architecture behind the assistant: roughly 64 purpose-built tools, an approval-gated write system that never lets the AI act without human sign-off, and the logging system Josh built to audit a codebase that, by his own admission, has become a black box even to him.\nJosh walks us through his full build process, from sketching UI mockups in Adobe Illustrator to having Claude draft implementation plans and Codex tear them apart before anything ships, plus how he red-teams his own assistant before any change touches its tools or system prompt. He also shares how Elephants evolved from a low-code personal relationship manager built on Glide Apps and Make.com into a full-stack platform now running the daily operations of NC State's Entrepreneurship Clinic.\nIt's a candid look at what it actually takes to ship an AI feature people trust with real operational data, including the parts that still keep a solo founder up at night.\nTimestamps\n00:00 Cold open: \"The devil's in the details\"\n00:45 Welcome back to Redacted\n01:20 Introducing guest Josh Guter\n02:33 Josh joins and how he and David met\n02:56 What Elephants is: ops software for small teams\n07:03 The real product: the AI assistant\n08:04 The big question: model, cost, context, trust\n14:04 Behind the scenes: Anthropic's API and the Make.com origin story\n15:12 64 tools and counting\n16:01 Managing the internal prompt library\n18:07 Rebuilding full-stack with Opus 4.8 and Codex\n20:38 Auditing a black box: breadcrumb logging\n21:57 Claude Code as the new Make.com\n25:10 The build process, start to finish\n26:23 Illustrator mockups and the Claude/Codex review loop\n29:05 QA without a QA team: red-teaming the assistant\n31:51 Taylor's tangent: rebuilding an interview ingestion pipeline\n34:09 The founders who never read their own PRs\n35:02 Why nothing ships without user approval...","thumbnail_url":"https://img.transistorcdn.com/UEzoK7N1siD9YRaSVBWfZ8B3suSS2aonEIZ5NwBH1Gs/rs:fill:0:0:1/w:400/h:400/q:60/mb:500000/aHR0cHM6Ly9pbWct/dXBsb2FkLXByb2R1/Y3Rpb24udHJhbnNp/c3Rvci5mbS80NjBh/YWVhZTA4ZTUyY2Fl/MDdhNzQwZWFhNDI4/MDc3Zi5wbmc.webp","thumbnail_width":300,"thumbnail_height":300}