Hampshire Property Group: two days, 20 people, 10 working Claude tools

Where they started

HPG had already rolled out Claude on the Team plan, and its AI and digital platforms team was building internal tools of its own. What the wider business lacked was a shared way of working. When do you use a project, a skill or Cowork? How do you check what Claude gives you? And when should a good idea go to the AI team instead of becoming one more app that somebody built on their own?

Marketing had the sharpest version of the problem. Every campaign launch meant writing copy for each park's social pages, and with more than 100 of them, that took the team days.

Day one: how it works, and how to check it

We started with how the models actually work. People who understand context windows and retrieval write better prompts, and they trust outputs for the right reasons.

Then we got them to catch Claude out. Teams worked through a mock launch plan and business case that we had seeded with contradictions and gaps, in three passes: a blind read, a set of directed questions, then a verification pass where every answer had to be tied to an exact sentence. At one point Claude confidently described a New Zealand benefit structure that the document never mentioned. The rule the room took away: if you can't point to the sentence, the answer isn't supported, however good it looks.

The afternoon covered custom instructions, Projects, Skills and Cowork. The group connected Microsoft 365 (read-only), learned to link SharePoint folders so Cowork could work from live files, and set Cowork to manual approval while they found their feet. Opus 5.5 had been released the night before, so this group was among the first to use it.

Day two: build it like software

Day two treated a skill as a small software project. Define the problem, write user stories and acceptance criteria, design the inputs and outputs, add a check step before anything gets processed, test against the criteria, then version and release it.

Some participants had built skills overnight and shared them. One was a competitor ad tracker that scans Meta for competitors' ads and reports back; the person who built it estimates it saves 2 to 3 hours a week.

Then came the hackathon. Teams of three, with a builder, a navigator and a tester, had 90 minutes to build, ten minutes to swap and red-team another team's tool, and five minutes to present. One rule applied to everything: anything sent outside HPG, published, paid for or deleted needs a person to approve it.

The marketing team's Cowork workflow

The marketing team set up a Claude project that takes in a campaign brief. Skills running in Cowork write the copy for each park page and lay it out for review. A marketing manager works through the drafts, approving or commenting. Comments go back in, Claude rewrites, and approved posts move on to the scheduling tool for the people who run each channel.

The approval step is deliberate. HPG's brand runs across more than 100 pages, and a person still signs off every post before it's published.

Everything else they built

In the same 90 minutes, other teams built:

• a guided report builder over the reservations system, where staff ask a question in plain English and get a PowerPoint or CSV back
• a utilities contract tracker with renewal dates and a "needs attention" list
• a headcount allocation tool for monthly hiring requests
• a site inspection tool for acquisitions that scores a park and produces a PowerPoint report
• an IT support scorecard built from ticket data
• an executive trading dashboard showing pace against forecast and the best and worst performing parks
• a park manager KPI dashboard with separate views for park managers, operations managers and leadership
• a graduate program tracker covering intake, interviews, offers and rotations
• a PMO health check connected to Asana

Keeping it under control

Ten tools in an afternoon brings its own risk: a business full of one-off apps that someone eventually has to support. We closed on a simple filter, "should you" before "could you", and on adding features to existing departmental tools before starting new ones. Anything that needs proper authentication, a real database or calls to outside systems goes to HPG's AI team to be made production-ready.

HPG is setting up an AI hub to match. Each function, starting with marketing and HR, forms its own group that meets monthly and feeds ideas back to the central AI team. The full cohort will come back together every six months to share what's working.