/architect
/architect is the advisory skill. Before you scaffold anything, it works out
what to build: from what you're building, what your data is, and what it's for, it
recommends a concrete architecture across six slots — storage, catalog, compute, access,
hosting, and metadata — then hands off to the build skills like
/new-portal. It decides; it doesn't build.
It's interactive: if your brief is thin it interviews you in short rounds, echoes an architecture brief back for confirmation, and never dead-ends on missing input — every question has a sensible default. See the architecture decision framework for the full model it encodes.
When to use it
Run it first, when you're unsure how to set up a portal — especially the data
infrastructure underneath it (files vs Git-LFS vs a lakehouse vs a warehouse; static vs
runtime; which metadata standard). If you already know what you want, skip straight to
/new-portal.
Inputs
| Input | Required | Notes |
|---|---|---|
| What you're building | No | Kind of portal + one or many publishers. Asked if missing. |
| Your data | No | Size, shape, update cadence, count, sensitivity. Point at files and it inspects them. |
| What it's for | No | Publishing, analytics, redistribution/harvesting, compliance. |
| Constraints | No | Team/budget, cloud preference (default Cloudflare), existing backend. |
Nothing is required — with no input it runs the four-round interview.
What it produces
- An architecture brief filling six slots, with one-line reasoning per slot and any
deviations from the opinionated default called out. The default stack is
git + giftless/R2 + Parquet + DuckLake + DuckDB, static on Cloudflare Pages, with Frictionless metadata — kept S3-compatible so R2 is never a hard lock-in. - An
ARCHITECTURE.mdwritten to your project documenting the decision. - A handoff: the exact next commands (
/new-portalwith the inferred namespace mode,/add-dataset,/connect-ckan,/deploy, …), and an offer to run the first one.
Example
/architect We're a national statistics office. ~200 datasets, mostly large CSVs
(some GBs), updated quarterly, all public, and we must publish DCAT-AP for the EU
data portal.
It infers a multi-publisher open-data portal with analytics-grade data and harvesting
needs, and recommends Parquet on R2 + DuckLake + DuckDB, static on Cloudflare
Pages, Frictionless + DCAT-AP metadata, and an owner namespace — flagging the
DCAT export and Git-LFS ingest as designed-in but built later — then hands off to
/new-portal.