Community Catchup #3

The third weekly community catchup brought together 64 members for 70 minutes of AI news, live demos, and the kind of conversations you only get when a room full of people are all building things at pace.

Highlights

The Auckland in-person meetup on April 8th was a hit — big turnout, great conversations, and a clear need for a bigger venue next time. The community is growing faster than the spaces can hold it.

The News Cycle

Adam ran through a rapid-fire AI news roundup pulled from the WhatsApp community threads. The headline: Anthropic’s new model benchmarks were remarkable — particularly its ability to find security exploits. Slightly worrying, but Anthropic are handling it well by giving early access to major organisations to patch vulnerabilities before wider release. OpenAI extended double usage limits to end of May. Meta, GLM, and Gemma 4 are all closing in on the frontier fast — Gemma 4 running locally on an iPhone being the most striking proof point.

The Axios npm supply chain attack hit close to home for anyone who’s been writing JavaScript for more than five minutes. One of the most-used packages in the ecosystem, compromised. A good reminder that giving agents broad access to install packages isn’t without risk.

Jake’s Travel Agent

Jake built Tucky — a modern TripIt-style travel itinerary manager — in three to four days, entirely in Claude Code. Not one line of code written by hand. Forward a booking confirmation email and an agent processes the PDF, works out which trip it belongs to, adds the details, and replies to confirm. Architecture spans MongoDB Atlas, Vercel, Cloudflare, Gemini Flash, Resend, and LangSmith — Claude suggested Gemini 2.5 Flash over 3.1 because 2.5 was cheaper and had a bigger free tier.

When asked how specific your prompts need to be, Jake was clear: domain expertise matters. His background as a technical architect meant he drove the data model, API design, and security model deliberately. “Claude’s really just like a collaborator — it’s like I’m on a whiteboard with another colleague, and they go and do all the work as well.”

An Events App That Actually Works

Another community member demoed a social events coordination tool they’d launched that week — built to solve the genuine problem of herding friends to actually book tickets in advance. Scrapes events, sends invites to your groups, handles both general admission and allocated seating flows, tracks who’s in. Independent theatres are already reaching out.

Claire’s Agent Architecture

Claire brought a deep dive into how she rebuilt her AI consultancy’s entire operating system. Notion as the knowledge hub, Claude Code as the agent layer, with a structured map so agents always know how to navigate client information. The standout piece: a CIO skill that evaluates incoming LinkedIn posts against her clients’ agent stacks — filtering genuine alpha from noise, and when something is worth pursuing, proceeding to implement it.

Her non-technical fractional team? They use Notion AI directly. No Claude subscription, no terminal, no friction. It just works. “The delta of how much work I’m having to do versus how much is just coming out of the box has shrunk completely.”

NZ Open Data Skills Repository

The most concrete idea to come out of the session: a community GitHub skills repository full of NZ-specific data connectors — LINZ, Stats NZ, Auckland Transport, weather — that anyone can point an agent at.

The analogy that landed: think Folding@home, but for spare AI tokens. Instead of donating idle CPU cycles to protein folding, the community contributes skills that make NZ’s public data actually usable by agents. Peter already has working skills for weather data and Auckland Transport. Simon flagged census data as a high-value baseline that most NZ projects could benefit from. Jason raised a practical blocker — a lot of government resources are locked up in PDFs and formats agents can’t easily consume — and suggested converting them to markdown-ready formats as a first contribution.

The repo is live: github.com/thecolab-ai/.skills. If you’re building on NZ data, this is the place to contribute.

What’s Coming

Adam also outlined a Claude Impact Lab — an Anthropic-partnered hackathon focused on using AI to solve real societal problems. Likely Auckland-based, with prizes and Anthropic involvement in judging. A Wellington edition is also being explored.

On the government front: direct outreach works. A message to Minister Collins led to multiple meetings with senior digital government figures within two weeks. The appetite is there. The community has the tools. Watch this space.