Vibe Coding & Web App Development with AI

Greg Dickson opened the evening with a practical case for turning static internal documentation into usable software. His premise was simple: most organisations already have the raw material for internal tools sitting in Google Docs, spreadsheets, and standard operating procedures — but those processes are rarely followed consistently because they remain boring, static, and hard to use.

Instead of waiting on internal dev teams or paying agencies to build bespoke systems, Greg showed how non-developers can use Claude as a product manager and Manis as the engineering layer. His workflow was clear and replicable: dump the SOP into Claude, have it summarise the workflow and write a proper spec, then hand that spec over to Manis to generate the application.

The live demo made the point. Starting from a YouTube best-practices article, Greg used Claude to create a detailed product requirements document, then used Manis to turn that into a functioning internal checklist app in under half an hour. The result included review workflows, performance tracking, team roles, and dummy data — more than enough to prove the concept.

Greg was refreshingly honest about the trade-offs. This was not a sermon about best practice. In his words, the code may break rules, but it gets the job done. For internal tools, MVPs, and quick experiments, that speed matters. His strongest point was that when non-developers can prototype their own ideas, they reduce bottlenecks, create better buy-in, and learn how to communicate specifications more clearly to technical teams later.

The most useful takeaway from Greg's talk was not just that these tools are fast, but that play is now a productive business activity. Throw a process into the machine, see what comes back, refine it, and learn where the boundaries really are.

Caelan Huntress took the idea further, framing vibe coding not just as a software technique, but as a creative shift. Using Rick Rubin as the patron saint of the movement, Caelan argued that the human role in AI-assisted building is not to type every line, but to bring taste, direction, and decisiveness. AI can generate endless options; humans still decide what is worth making.

The talk moved fluidly between philosophy and live demos. Caelan walked the room through a maturity ladder of AI building — from plain prompting, to custom GPTs and Gems, to skills, Claude co-work, and finally Claude Code connected to GitHub. This helped the audience place themselves on the map, whether they were complete beginners or already experimenting with agentic tooling.

The live demo was classic vibe coding: the audience named a new Christchurch cafe, Caelan generated a logo and dummy menu data in parallel, then used Claude Code to build a mobile-friendly one-page website. At the same time, Puck — his OpenClaw agent — attempted the same task in parallel on a live deployment. The local version worked quickly; the public deployment hit friction. That contrast was actually useful. It showed the room where AI feels magical and where real-world deployment still demands oversight.

Caelan also used the recent accidental Claude Code code release as a teachable moment, surfacing what operators can learn from the architecture beneath the interface: lean memory files, self-improvement loops, spinner verbs, and why reasoning structure matters. The message was that the most valuable AI operators won't just prompt well — they'll understand enough of the machinery to direct, evaluate, and improve it.

One of the most memorable moments was the juggling metaphor. Managing agents, Caelan argued, is like keeping multiple balls in the air. You drop things. Everyone who tries this seriously drops things. But the people trying, learning, and recovering are also the people building new capability. In that frame, mistakes aren't evidence of failure; they're part of becoming effective.