Community Catchup #11

The eleventh community catchup was a practical deep dive into what happens after the novelty wears off and teams start using AI as a serious production tool: model selection, documentation, review loops, compliance, and business models.

Highlights

The recurring phrase was domain knowledge is still required. The tools are dramatically faster, but the best results came from people who knew what outcome they wanted, supplied the right context, and built guardrails around the work.

The New Model Mix

The session opened with the now-standard Friday question: who has any tokens left? That turned into a useful conversation about using the right model for the job. Premium models make sense for hard reasoning and architecture. Cheaper models can handle research, web search, and routine subtasks. Local models are becoming relevant for specific workflows where privacy, cost, or control matter.

Adam shared observations from Japan: model releases are accelerating, system prompts are shrinking, and the frontier labs are increasingly telling people to get out of the model’s way. Steve framed it as different tiers of agentic engineering — frontier-lab token abundance, practitioner budgets, and low-cost open-source workflows all require different tactics.

Jason’s 36-Hour Website Rebuild

Jason rebuilt his company website in 36 hours. The important detail: it was not magic prompting. He had to give Claude the right business context, product knowledge, Microsoft 365 and Graph capability, and clear rules around factual claims. Without that, it drifted toward a generic IT website. With the right context, it produced something usable fast.

That led into one of the strongest operating lessons of the session: document what the agent learns. Jason’s team makes the AI write down how it deployed systems, what patterns it followed, what compliance checks it ran, and how to reproduce the work. Thinking is expensive; referencing documented patterns is cheap.

Commercial Readiness as a Product

Jason then demoed a commercial readiness platform for software projects. Link a GitHub repo, run an assessment, and it scores the project across architecture, security, compliance, operations, documentation, release governance, payment correctness, and more. It can compare outputs across models, produce executive summaries, create remediation workbooks, and generate prompts or scripts to fix the problems.

The pitch was simple: if New Zealand is going to build more software, it needs to build software that is secure, scalable, and investment-ready. The platform is intended to be available to the community with bring-your-own-key access.

Products That Would Not Have Happened

A theme emerged around projects that were previously too expensive, too slow, or too stalled to justify. Claire rebuilt an app stack in three weeks that had once taken six months and $150k. Connor’s Propel Music platform moved from an eight-year idea into a live product with hundreds of DJs. Jason’s website rebuild happened in days.

Steve suggested documenting these stories because they are a useful counterweight to the doom narrative. The point is not just that AI makes existing work faster; it makes new work economically possible.

Toward Builders Workshops

The group closed around workshops and shared practice. Advanced sessions, AWS spaces, hands-on builders events, and more examples of goals, prompts, and review loops all came up. The community is clearly shifting from “look what I made” toward “here is the repeatable way I made it.”