Hosting by workload, not by label

Demo kit for the talk "Hosting by workload, not by label: choosing cloud for education data and AI" at the Giga Connectivity Forum 2026 in Geneva.

Two education services carry the same label, "education data in the cloud". One holds student records and exam results, the other is an AI assistant for teachers. In the demo, Claude Code works out what each service needs from its hosting with the GIGA cloud skills, and shows how one platform on IaaS and Kubernetes can run both. The results appear in a small web app.

Start state online: https://hosting-by-workload.vanillacore.net

What is in the kit

Path	Content
demo/	The folder in which Claude Code runs: CLAUDE.md (frame for the web app, colours, writing rules) and the two cases in cases/. The web app is built into demo/out/.
app/assets/	Look of the web app: stylesheet, fonts, Giga logo.
scripts/	Build, reset, serve and rehearse.
marketplace/	Claude Code marketplace entry for the giga-cloud-skills plugin.
deploy/	Kubernetes manifests for the public start state.

demo/ holds only what the agent should see. CLAUDE.md sets the frame, not the content: the agent works out every result live with the skills.

Run the demo

Requirements: Python 3 and Claude Code with access to a Claude model.

1. Install the GIGA cloud skills as a Claude Code plugin:

   claude plugin marketplace add ./marketplace
   claude plugin install giga-cloud-skills@giga --scope user
   

   Other routes, including Codex, are described in Install and use.

2. Build the web app in its start state and serve it:

   scripts/reset.sh
   scripts/serve.sh            # http://localhost:8765/
   

3. Start a new Claude Code session in demo/ and send the two prompts one after the other:

   cd demo && claude
   

   Look at the two cases in cases/. What does each one need from its hosting, and how can one platform run both? Use the GIGA cloud skills and show it on the planning page.
   

   Plan that platform with IaaS and Kubernetes and show it on the platform page, with an architecture diagram.
   

   The first prompt fills 1 · Planning, the second 2 · Platform. Both pages reload every 5 seconds until the agent has written them. Each prompt takes about two to four minutes.

Run scripts/reset.sh before every new attempt and start a fresh session afterwards, not claude -c. Use only one Claude Code session in demo/ at a time, because two sessions write the same pages.

Rehearse

scripts/rehearse.sh resets the app and runs both prompts in one continued, non-interactive session. Pages, logs and timings go to rehearsals/<time>/. To see which skills ran:

python3 scripts/show-tools.py rehearsals/<time>/*.jsonl

Results differ from run to run in layout and wording. The skill calls stayed the same in rehearsal: workload-cloud-fit-assessment and public-service-platform-architecture for the first prompt, iaas-foundation-practitioner and kubernetes-cluster-lifecycle-practitioner for the second.

After the demo

• Cloud for DPI on GitBook explains the route from workload to platform in depth.
• The management dashboard rolls out and manages clusters and deploys workloads through a graphical interface. It implements what the chapters teach.

Public start state

deploy/ serves the start state at https://hosting-by-workload.vanillacore.net: start page, use cases and the two result pages before the agent has filled them. Rebuild the static files after a change to the cases or the look:

python3 scripts/build-app.py --public --out deploy/site
kubectl kustomize deploy

Licences

Fonts and logo are third-party material, see NOTICE.md.