PortalJS — Architect
Overview
The advisory entry point for a PortalJS project. Before anything gets scaffolded, this
skill works out what to build: given the kind of portal, the shape of the data, and its
purpose, it fills six architecture slots (storage, catalog, compute, access, hosting,
metadata), resolves two build-time knobs (per-dataset data tier and the portal-wide
DATA_QUERY mode), and hands off to the concrete build skills. It decides; it does not
build. When the brief is thin it interviews in short rounds and never dead-ends — every
question has a sensible default, reachable by replying "use defaults."
Prerequisites
- A rough idea of the portal's purpose and data (exact numbers are not required — the interview supplies defaults for anything missing).
- Optional: local files or a directory of sample data to inspect for size and shape.
- No PortalJS project needs to exist yet; this skill runs before scaffolding.
Instructions
The canonical, full step-by-step workflow lives in
.claude/commands/portaljs-architect.md —
that file is the single source of truth. Follow it when executing this skill:
- Parse
$ARGUMENTSfor anything already specified, then interview for what's missing, one round at a time: (1) what's being built, (2) what the data is, (3) what it's for, (4) constraints. Accept "use defaults" at any point. Inspect named files/directories withdu -shand line counts to ground size guesses. - Derive the recommendation by matching the answers against the decision tables —
Storage/Catalog/Compute by data volume and query needs, Access/Hosting by public
vs. private, Metadata by standards-compliance needs — then resolve the two
build-time knobs: per-dataset data tier (
inline | LFS | external) and the portal-wideDATA_QUERYmode (flat | duckdb). - Echo the architecture brief (stack, reasoning per slot, deviations from default, deferred items) and wait for confirmation ("go") or corrections.
- On confirmation, persist the brief to
./ARCHITECTURE.mdin the working directory. - Hand off to the build skills —
/portaljs-new-portal,/portaljs-add-dataset,/portaljs-connect-ckan,/portaljs-define-schema,/portaljs-deploy— mapped from the brief, and offer to run the first one.
Output
- Created:
./ARCHITECTURE.mddocumenting the six slots, the two build-time knobs, the reasoning, and anything deferred to a later build step. - Modified: nothing else — this skill is advisory only.
- Verified: the brief was echoed back and confirmed before being persisted.
- Result: a concrete, named sequence of follow-up skill invocations
(e.g.
/portaljs-new-portal→/portaljs-add-dataset→/portaljs-deploy).
Error Handling
| Symptom | Cause | Fix |
|---|---|---|
| Skill keeps asking rounds of questions | Brief was thin or $ARGUMENTS omitted | Answer inline, or reply "use defaults" to accept the opinionated default stack |
| Recommendation looks generic | Rounds were skipped without real data details | Give actual size/shape/cadence, or point at files for du -sh inspection |
ARCHITECTURE.md never appears | Confirmation step was skipped | Reply "go" once the echoed brief looks right |
Scaffolded portal has the wrong DATA_QUERY | Flat downgrade wasn't applied | Run the perl -pi -e one-liner from the command file against lib/datasets.ts |
| Hand-off names a skill that doesn't exist | Decision maps to a (planned) skill (e.g. /connect-openmetadata) | Treat it as designed-in/built-later; proceed with the closest available skill |
Examples
Example 1 — National statistics office, DCAT-AP harvesting
/portaljs-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.
Infers a multi-publisher, analytics-grade portal. Recommends Parquet on R2 + DuckLake +
DuckDB, static Cloudflare Pages, Frictionless + DCAT-AP metadata, owner namespace,
data tier external for the Parquet, DATA_QUERY=duckdb. Writes ARCHITECTURE.md and
hands off to /portaljs-new-portal then /portaljs-add-dataset.
Example 2 — Small nonprofit, no arguments given
/portaljs-architect
Runs the full four-round interview since nothing was pre-filled. Accepting defaults at
each round lands on the opinionated default stack: repo files or Git-LFS + R2 storage,
datasets.json catalog, DuckDB compute, static access on Cloudflare Pages, Frictionless
metadata, theme namespace, data tier LFS, DATA_QUERY=duckdb.
Example 3 — Internal catalog with restricted datasets
/portaljs-architect Internal engineering data catalog, single team, dozens of CSVs, some of it access-controlled to specific roles.
The private-data answer in Round 2 flips Access/Hosting to runtime + backend RBAC on Cloudflare Workers — flagged as the larger, opt-in build — while Storage/Catalog/Compute still follow the volume-based defaults.
Resources
- Full workflow:
.claude/commands/portaljs-architect.md - Reference:
references/reference.md - Decision framework:
site/content/docs/architecture/decision-framework.md - Related skills:
/portaljs-new-portal,/portaljs-add-dataset,/portaljs-connect-ckan,/portaljs-define-schema,/portaljs-deploy
