Role-based training curriculum
One umbrella curriculum with modules per role — not ten separate courses. Every role's job changes when agents do the typing, and this curriculum is how each role learns its new one. It is deliberately practical: every module ends in a lab against the RAPID project scaffold with a graded exercise, because directing and reviewing agents are skills acquired by doing, not by watching slides.
The shape
- One curriculum, role-based modules — a common foundation everyone takes, plus the modules a role needs. Materials live in /toolkit/training.
- Labs, not lectures — every module runs hands-on against the scaffold: same
CLAUDE.md, memory and gates as a real product. - "Running STOP with agents" ships first — STOP is the safe training ground for the direct → review → govern muscle: real remediation work where mistakes are recoverable.
- Foundation before role modules — nobody directs an agent before completing the responsible-AI foundation.
The modules
| Module | Audience | Objectives | Lab format |
|---|---|---|---|
| Foundation: responsible AI use & AI literacy | Everyone — product, engineering, test, design, SMEs, Support, Commercial, leadership | EU AI Act Article 4 AI literacy; where agent autonomy ends (the irreversible list); data handling — no client or policyholder data in a prompt, UK GDPR applies to a context window like anywhere else | Scenario exercise: classify agent actions as proceed / stop-and-authorise; graded |
| Running STOP with agents (delivered first) | Whole squads, UK + India | The direct → review → govern loop, learned on legacy remediation | Run code-archaeology against a fixture legacy repo; graded knowledge-pack review |
| Directing agents | Senior + Intermediate Devs | Decompose work into agent-sized briefs; write directing prompts; review at checkpoints; know when to take the keyboard back (build loop) | Build a small feature in the scaffold by direction only; graded on first-pass acceptance and merge-gate pass |
| Spec-writing for agents | PO / BA | Turn a PFD into a spec an agent can build from unaided — testable acceptance criteria, explicit constraints (the PO directing agents) | Write a spec from a sample MTA requirement; an agent builds from it; graded on how far it gets without a clarifying question |
| Rules as decision tables | SMEs | Express underwriting and claims rules as decision tables and worked examples an agent can consume (SME engagement) | Convert a rules narrative into a decision table; an agent implements it; graded on coverage of seeded edge cases |
| Reviewing AI code and tests | Devs / Testers | The characteristic failure modes of generated code and generated tests — plausible-but-wrong logic, happy-path bias, assertion-free tests | Seeded-defect review in the scaffold; graded on defects found vs seeded |
| Design-agent direction | Design | Direct design agents against the design-phase artefacts and the insurance UX heuristics | Produce a screen spec via a design agent, then crit it; graded crit |
| Security failure modes of generated code | Devs / Testers / Sec | The vulnerability classes agents most often introduce — injection, authorisation gaps, secrets in code, plus LLM-app risks | Find seeded vulnerabilities in a generated scaffold service; graded findings list |
| Support handover | Support | Operating a weekly-release product: runbooks, dashboards, the rapid-fix path, AI-defect incidents — per the support playbook | Triage a simulated incident end-to-end in the scaffold; graded triage decisions |
Labs & grading
- Graded means human-reviewed — pass or redo with feedback, against a published pass criterion per lab. Grades exist to certify readiness, never to rank people (the metrics anti-gaming rule applies).
- Completion recorded per person — it satisfies the week-1 module requirement in squad onboarding, provides the Article 4 AI-literacy evidence, and feeds the adoption view of the metrics platform.
- Refreshed with the toolkit — when a prompt library or guardrail changes materially, the affected module and its lab change with it.
Pending — authoring & scheduling
The module list and lab formats above are the agreed scope; module authoring (a named author per module, drawn from the shared specialists and the CoP) and the UK/India delivery calendar around the overlap window are still to be assigned.
Standards referenced: EU AI Act Article 4.