Your job
Keep teams moving, keep it inclusive, and ensure each group leaves with:
- A demo/prototype
- A clear housing outcome
- Assumptions + risks/guardrails + optional next steps written down
Non-negotiables to set early
- Tenant respect: ensure lived experience is heard and not tokenised.
- Data safety: no personal data or case notes in AI tools; use synthetic/anonymised examples.
- Small PoC mindset: one thin user journey that works beats a big idea that doesn’t.
- Show your working: assumptions, risks, and next steps must be captured in a short README/doc.
Suggested running order
09:00–09:20 Welcome, rules, problem statement, example PoC bar (“small and demo-able”).
09:20–09:45 Team formation + idea selection (impact/effort filter). Ensure tenants distributed across teams.
09:45–12:00 Build sprint. Roam, unblock, keep scope tight.
12:00-12:30 Prepare demo + ensure 1-page README exists (assumptions, data needed, risks, next step)
12:30–13:15 Working lunch.
13:15–14:00 Demos (2–3 mins each), wrap, contacts, capture follow-ups (who wants to continue collaborating).
Parallelise aggressively
If you see lots of people around one screen, intervene:
- “Split into parallel roles for 15–20 minutes, then reconverge.”
Recommended roles (fast + inclusive)
- Driver: owns the laptop and implements.
- Navigator: decides next step, keeps scope tight, reviews AI output.
- Tenant advocate: checks the workflow matches real needs and avoids unintended harm.
- Prompt lead: runs structured prompts (plan → small diffs → tests → risk checks).
- Scribe: writes README live (assumptions, risks, decisions, next steps).
- Pitch lead: crafts the 2–3 minute story and demo flow.
How to split work in practice (what you say to teams)
- “One person: prompts for the plan + acceptance criteria.”
- “One person: prompts for UI copy + the user journey.”
- “One person: prompts for risks/guardrails (privacy, bias, safeguarding).”
- “Driver: builds the thinnest working slice.” (potentially have two distinct of these)
- “Scribe: captures decisions continuously.”
Unblocking patterns (what to do when teams stall)
- They’re stuck in “idea soup”
Say: “What is the user doing in minute 1?”
Then force a one-screen PoC: input → transform → output. - AI outputs are wandering or low quality
Make them add:
• Output format: “Return JSON/table/checklist/only code diff.”
• Constraints: “Max 2 screens; works without real data; 90-minute build.”
• Examples: provide 1–3 synthetic records and desired outputs. - They’re generating big opaque code dumps
Say: “Smallest working version first. Then add one feature. Output only the changes each step.” - Tenant voices aren’t being used
Pause for 5 minutes and ask tenants:
• “What would make this useless?”
• “What would make this unsafe?”
Capture answers in the README and adjust scope accordingly.
Facilitator prompt cheat sheet (give these to teams)
PoC definition
“Define a PoC we can demo in 90 minutes. Include: target user, steps, synthetic data, outputs, and what we are not doing.”
Acceptance criteria
“Write 6 acceptance criteria + 3 edge cases + what failure looks like.”
Risk/guardrails
“List top 10 risks (privacy, bias, hallucination, safeguarding, misuse). Provide mitigations and what to disclose to users.”
Demo script
“Write a 2-minute demo script: problem → user story → demo steps → impact → next step.”
License
This page is licensed under CC BY-SA 4.0 (share/adapt with attribution and the same license): https://creativecommons.org/licenses/by-sa/4.0/
HAILIE 