Community Catchup #10

The tenth community catchup brought together the regular Friday crew for a wide-ranging session on model reliability, token costs, reusable skills, business adoption, and the growing need to turn AI experiments into repeatable infrastructure.

Highlights

The big theme was moving from clever demos to durable systems. Models are getting stronger, but the community kept circling back to the same practical questions: how do you keep costs under control, how do you make data accessible safely, and how do you help real organisations adopt this without fear?

When the Model Moves Under You

Peter kicked things off with a warning from the trenches: Copilot had drifted enough that daily email workflows started inventing names. The group treated it as a reminder that AI automations are not static software. If the model changes, your workflow can change too — even if your code did not.

That led naturally into token economics. Adam’s live website build had used 51 million tokens, with the vast majority coming from cached reads. Steve’s question landed well: does investing more effort up front in a tight plan reduce the expensive back-and-forth later? Adam’s answer was practical — use plans, handoff docs, and smaller delegated tasks so each agent does less rediscovery.

Skills as the New Interface

Adam walked through the community skills repository: reusable connectors and instructions for public New Zealand data. Child poverty, MSD benefits, groceries, bin schedules, Auckland and Wellington services — the idea is that agents should not have to rediscover the same APIs every time.

The most concrete demo was refreshingly mundane: an Auckland bin schedule skill told Adam his collection day had shifted after a public holiday. Small, boring, useful. Exactly the kind of thing that shows why local skills matter.

Building for Business Reality

Jason showed Ballad, an AI-built learning platform designed around multiple ways people absorb information: read, listen, speed-read, skim, recall. The platform was built with Claude, but the more interesting point was the product thinking behind it — open information can become practical education, and Claude is now good enough to build serious scaffolding quickly when the direction is clear.

The group also dug into authentication, data warehouses, Xero APIs, and AI usage monitoring. The shared conclusion: agents need structured access to data. Throwing huge JSON dumps into context is brittle and expensive; putting data into a queryable warehouse or well-designed skill is far more useful.

Adoption Is a Human Problem

Miriam brought a different lens: habit change. Her point was simple and important — fear shuts down people’s ability to learn new behaviours. If AI is introduced as a cost-cutting threat, adoption will stall or even be sabotaged. If it is introduced as augmentation and innovation, people are much more likely to engage.

That reframed the week’s technical discussion. The tools are accelerating, but the bottleneck is often trust, communication, and leadership. For New Zealand to benefit, the community needs to help organisations move past fear and toward useful, human-centred adoption.

What Comes Next

The skills repo will keep growing, Jason plans to share more of his commercial readiness and learning platform work, Adrian is looking for a data warehouse proof-of-concept partner, and Adam is continuing to shape the Impact Labs idea around real public-good problems backed by usable data.