Tableau init

tableau-init

The first step of the tableau-dashboard-plugin workflow. Running it in a project directory scaffolds a complete starting skeleton and writes the STATE.md manifest that every downstream skill reads to know where the analyst is.

This skill is step 1 of 8. It has no upstream dependencies — it is what creates the workflow state in the first place.

Reads	— (nothing; this step bootstraps the project)
Writes	STATE.md + a scaffold/ folder of demo examples: scaffold/EXAMPLE-DASHBOARD-REQUEST.md, scaffold/EXAMPLE-datasources.json, scaffold/.env.example, scaffold/branding/EXAMPLE-branding.md, scaffold/sample-data/*.csv
STATE.md update	Creates the manifest with init = approved and steps 2–8 = pending; records target_tableau_version, data_mode: csv, current_version: v_1
Next step	tableau-intake (or tableau-data if the analyst skips intake)

scaffold/ examples vs. production inputs

Everything init writes under scaffold/ is a demo example so the analyst can trial the whole workflow before owning any real inputs. Every downstream skill prefers the production input at the project root and only falls back to the scaffold/ example when the production file is absent:

production input (preferred)	scaffold/ demo fallback
DASHBOARD-REQUEST.md	scaffold/EXAMPLE-DASHBOARD-REQUEST.md
datasources.json + .env (published-ds route)	scaffold/EXAMPLE-datasources.json + scaffold/.env.example
branding/branding.md	scaffold/branding/EXAMPLE-branding.md
data/*.csv (csv route)	scaffold/sample-data/*.csv

init deliberately creates only the scaffold/ examples (and STATE.md), never the production files — their absence is exactly what tells the analyst (and the model) which real inputs are still owed. See CONTRACT.md §3 for the canonical rule.

How to run

1. Determine the project directory. Default to the current working directory. This is where the skeleton and STATE.md will be written.

2. Ask for the target Tableau Desktop version — unless the user already stated it. This is recorded once, here, and drives tableau-build later (it is never re-asked downstream). Offer exactly the two allowed values:

   • 2024.2-2025.x — the broad-compatibility target (no rounded corners, no <explain-data> in the workbook).
   • 2026.1+ — the newest target (enables border-radius and newer XML).

   If unsure, recommend 2024.2-2025.x for the widest compatibility.

3. Run the scaffolder, passing the chosen version:

   python "${CLAUDE_SKILL_DIR}/scripts/init.py" --target-version "<version>" "<project-dir>"
   

   (Use python3 if python is unavailable.) The script copies this skill's skeleton/ templates into the project's scaffold/ folder and writes a schema-valid STATE.md. It is non-destructive: any file that already exists — including an existing STATE.md — is preserved, never overwritten, so re-running on an established project keeps the analyst's edits and pipeline progress.

4. Relay the scaffolder's summary to the analyst: what was created, what was preserved, and the one input to act on now — create their real DASHBOARD-REQUEST.md at the project root (copy scaffold/'s example, or just paste the request when running tableau-intake). Then tell them to open a fresh conversation and run the next step (tableau-intake to structure the request, or skip straight to tableau-data if they prefer). Suggest tableau-route if they want the router to confirm the next step.

   Don't walk the analyst through data acquisition or branding here. Those decisions belong to the skills that own them — tableau-data explains the two data routes (real CSVs in data/, or a published data source queried via the VizQL Data Service, with the scaffold/sample-data/ demo as fallback) when the analyst reaches that step, and tableau-brand handles branding. Init only lays the examples down; it does not pre-explain how to fill them.

Notes

• The scaffold/ examples (including scaffold/sample-data/*.csv) let the project run a full demo immediately. How the analyst swaps in real data — into data/ — is tableau-data's topic, not init's; don't describe acquisition modes here.
• target_tableau_version is the one decision captured at init time. Everything else (the request, the data, branding) is filled in by the later steps that own those concerns.

The full STATE.md schema and the ordering / staleness / versioning rules live in CONTRACT.md at the repo root. This skill restates only its own slice; init.py is the executable mirror of the schema it writes.