| name | healthcare |
| description | End-to-end provisioning of a healthcare host on Alvera, walking the canonical 11-phase setup defined by `tests/healthcare/_order.json` in `@alvera-ai/platform-sdk@0.8.0` (the executable contract). Each phase delegates to a primitive skill (`guided`, `custom-dataset-creation`, `DAC-upload`, `agentic-workflow-creation`, `query-datasets`) and ends with a verification pointer back at the corresponding integration-tests spec, so the conversational setup and the automated suite never drift apart. Use when the user says "set up healthcare", "provision a healthcare tenant", "walk me through healthcare setup end to end", or similar. Domain-first orchestrator — invoke the primitive skills directly for ad-hoc / out-of-sequence work.
|
Healthcare host setup
End-to-end domain orchestrator. The user describes their healthcare data;
this skill walks them through the same 11 phases the integration-tests
suite executes against the dev stack — but with their real data instead
of synthetic fixtures.
SDK pin
This skill targets @alvera-ai/platform-sdk@0.8.0 exactly. The walk
below mirrors the manifest at that tag:
https://github.com/alvera-ai/platform-sdk/blob/v0.8.0/integration-tests/tests/healthcare/_order.json
If the user has a different SDK version installed, prompt to install
@alvera-ai/platform-sdk@0.8.0 before continuing — version drift breaks
the phase-to-spec mapping below.
Prerequisites
alvera CLI reachable (delegate to /guided "Step 0" if not)- Active session for the target tenant (
alvera login, then alvera whoami
to verify) — the skill never sees the password
- A working dev / test environment of Alvera (this skill is not for prod
bring-up — provision in a sandbox tenant first)
The 11-phase walk
Each phase below is a single line: what gets provisioned + which
primitive skill drives it + the spec that verifies it landed.
| # | Phase | Primitive skill | Verifies via |
|---|
| 1 | bootstrap | /guided (datalake create) | bootstrap.test.ts |
| 2 | data-sources | /guided (data-sources create) | data-sources.test.ts |
| 3 | custom-datasets | /custom-dataset-creation | custom-datasets.test.ts |
| 4 | interoperability-contracts | /guided (interop-contracts create) | interoperability-contracts.test.ts |
| 5 | invite-team | /guided (admin invite) | invite-team.test.ts |
| 6 | tools | /guided (tools create) | tools.test.ts |
| 7 | create-dac | /guided (data-activation-clients create) | create-dac.test.ts |
| 8 | run-dac-single | /DAC-upload (single-row) | run-dac-single.test.ts |
| 9 | run-dac-bulk | /DAC-upload (CSV) | run-dac-bulk.test.ts |
| 10 | standard-workflow | /agentic-workflow-creation (template) | standard-workflow.test.ts |
| 11 | agent-driven-workflow | /agentic-workflow-creation (custom + AI agent) | agent-driven-workflow.test.ts |
Detailed walkthrough — what to elicit, what to confirm, what to verify —
lives in references/order-walk.md. Read it before opening the phase
loop.
Workflow
-
Confirm SDK pin — alvera --version should report 0.8.0. If not,
install: npm install -g @alvera-ai/platform-sdk@0.8.0 (or the user's
preferred package manager). Stop and ask if the user wants a different
pin — the phase-to-spec mapping is tied to this exact tag.
-
Open the phase loop — announce the 11-phase plan up front so the
user knows what they're agreeing to. Example:
"I'll walk you through the canonical healthcare host setup —
11 phases, end-to-end, mirroring the tests/healthcare/_order.json
sequence at platform-sdk@0.8.0. We can stop at any phase, skip
non-essential ones (invite-team is the obvious one), or pause and
resume later. Ready?"
-
Per phase (1 → 11):
- Read the corresponding section of
references/order-walk.md.
- Hand off to the primitive skill with a tight scoping handover —
don't reimplement what
/guided or /custom-dataset-creation
already does.
- On return, confirm the resource landed (one CLI list call) and
point at the integration-tests spec as the executable check the
user can run themselves:
pnpm test:healthcare -- <spec-name>.
- Append to
alvera-<tenant-slug>.yaml if the receipt is enabled.
-
Stop on first failure. If any phase fails (4xx, missing
prerequisite, validation error), do NOT skip to the next phase —
subsequent phases assume the previous one's output. Surface the
failure plainly, route the user to fix it (often by going back to
the relevant primitive skill in isolation), and pick up the walk
from the failed phase.
Stance: be a conductor, not a re-implementer
The five primitive skills already know how to provision their resources.
This skill's job is to:
- Sequence them correctly (per
_order.json)
- Hand off scoped context (datalake slug, prior phase IDs)
- Verify each phase before advancing
- Translate between the user's domain language ("our patient data",
"appointment reminders") and the primitive skills' resource language
Do NOT duplicate any elicitation logic the primitive skills own.
If the user wants to override a phase mid-walk (e.g. "I already have an
EMR data source — skip phase 2"), accept and verify by listing the
existing resource, then continue.
Out of scope
- Prod tenant setup — this skill is for dev / test environments.
Real prod onboarding has additional compliance gates (BAA, PHI handling
audits) outside this skill's purview.
- Cross-domain setup — for
accounts-receivable or payment-risk
industries, use those domain skills (currently stubs — they delegate
to /guided).
- Runtime ops — workflow execution, dataset search, and DAC ingest
beyond the canonical-walk verifications belong to runtime usage, not
setup. Out of scope here.
References
references/order-walk.md — per-phase elicitation + verificationreferences/fixtures.md — synthetic fixtures from platform-sdk and
how they relate to user-supplied real datareferences/example-transcript.md — sample 11-phase conversation
