Custom Dashboard Builder

Turns a list of metrics and filters — or an existing dashboard — into a single offline dashboard.html built from live Treasure Data queries, plus a dashboard skill for the project. Setup is a first draft with as few questions as possible; the user iterates afterwards through the dashboard skill (change metrics, filters, business context).

<skill_dir> = the folder containing this file — resolve it to an absolute path before running anything from it. Commands run from the Work Folder root. Templates in <skill_dir>/project_templates/ are read-only: copy, then edit the copy.

Treasure AI Studio questions: when the user asks how Treasure AI Studio works — Work Folders, permissions and access, sharing skills or agents, plugins and marketplaces, or where to find Studio documentation — answer from what you know, then share the official documentation: https://docs.treasure.ai/products/ai-studio/overview

Setup skill vs. project agent: this skill sets up, changes and rebuilds the solution; it is the same for every project. Once the project's own agent is installed in .claude/skills/ (<project_slug>-dashboard), questions about its results are answered by that agent. If this skill is invoked with such a question, read the installed agent's SKILL.md and answer by its rules (or tell the user to ask the agent in a new chat); if no agent is installed yet, say so and offer the setup.

The flow

Start with one sentence on what will happen and this table, then go straight into Step 1 — its first question doubles as the go-ahead. Use AskUserQuestion for every question, with a (Recommended) default; never ask what the data can answer. Between steps, give a 2–4 line summary and continue — don't wait for "go on".

Resume: if agent_outputs/*/project_state.md with Solution: dashboard-builder exists, ask "new dashboard or continue <project>?" and pick up at its Next Action.

Step 1 — Intake

Read references/intake.md. One AskUserQuestion batch:

• Starting point: "I have an existing dashboard" (attach a .dash export, an HTML file or a screenshot) / "I'll list the metrics and filters" / "Not now" (stop; nothing is written).
• Database: the database the tables live in — offer the likely ones from tdx databases (never assume).

Then read the dashboard (intake.md) or take the metric list, and derive the project name from the dashboard title (<project_slug>, lowercase with hyphens, e.g. sales-performance). Bootstrap the workspace — <skill_dir>/../shared/local_outputs.md Step 1 — and create agent_outputs/<project_slug>/project_state.md (<skill_dir>/../shared/project_state.md, Solution: dashboard-builder). Say once that all data in the HTML file is visible to anyone who opens it. Record the spec in project_state.md: Current phase "2 — data discovery", Next Action "Run data discovery".

Step 2 — Data Q&A and recommendation

Read references/data-discovery.md and run it without asking: tables and columns, the business time column and timezone, each metric validated as a formula on live data, joins checked for fan-out, the data-quality gate, dimension cardinality, payload estimate, query times. Ask only what the data can't answer (e.g. which column defines "cancelled", or a metric with two valid definitions).

Present one summary — metrics as formulas with their current values, filters, data range, caveats — and the recommendation (direct generate-data.js + template, or a workflow first) with its reasons. One question: accept, or switch path. Record everything in project_state.md → Discovery Findings, with the accepted path; Current phase "3 — workflow" or "4 — dashboard + skill" and the matching Next Action.

Step 3 — Workflow (only when recommended)

Read references/workflow-setup.md. Show the plan — SINK tables with their grain and measures, the refresh mode, and the defaults for the workflow project name and sink_database (checked to be free). One AskUserQuestion: "Build, run and validate (Recommended)" / "Change names first" / "Skip the workflow — build directly". On approval: copy and customize the template, push, run, monitor, validate the SINK tables against the source, and report (tables, rows, spot-checks, issues). If validation fails, fix and re-run before Step 4. Update project_state.md after the approval (names), after the run (attempt, result) and after validation (tables, checks) — Current phase "4 — dashboard + skill" once it passes.

Step 4 — First-draft dashboard and dashboard skill

Update project_state.md after each of the sub-steps below (Next Action = the next sub-step).

1. Dashboard — references/dashboard-build.md: copy project_templates/dashboard_template/ into agent_outputs/<project_slug>/dashboards/, write CONFIG, QUERIES and LAYOUT in generate-data.js, run it, spot-check ≥ 3 KPIs against the source (exact match), and show the dashboard.
2. Knowledge files and agent build — references/agent-setup.md: write business-context.md (inferences flagged for review), data-dictionary.md, metrics-catalog.md and sql-templates.md in agent_outputs/<project_slug>/, then build the skill in the staging folder (<skill_dir>/../shared/customer_agent_pattern.md Step 2).
3. Register — one AskUserQuestion: "Register as a skill in this Work Folder (Recommended)" / "Not yet". On yes: install to .claude/skills/<project_slug>-dashboard/ (pattern Step 3), run its build check, and show how to use it: "Show me the dashboard", "How is calculated?", "Add to the dashboard".
4. Update project_state.md: Current phase "complete — dashboard skill live" (or "awaiting registration"), what was built, open business-context items.

After Step 4 the user iterates through the dashboard skill. On request only: a 10-prompt evaluation of the skill (references/eval-setup.md + <skill_dir>/../shared/eval_pattern.md), a customer requirements questionnaire for the open business-context items (references/requirements_doc.md), handoff docs (references/handoff-docs.md), or a workflow change.

Rules

• Live data only. Every number comes from a query; data reaches the HTML only through generate-data.js. Never type data into files.
• Query results never enter the chat. Results flow tdx → generate-data.js → dashboard.html; checks use aggregate queries that return single numbers.
• Nothing changes in the account without approval — the workflow (Step 3) and the skill registration (Step 4) are the two approval points. Everything else is read-only.
• Local only. All files stay in the Work Folder (agent_outputs/<project_slug>/, .claude/skills/<project_slug>-dashboard/); nothing is uploaded or hosted. No generated dashboard.html or query output inside .claude/skills/.
• No PII rows in the HTML — aggregate, or mask.
• Templates and shipped scripts are pre-validated: fill them in, don't rewrite the engine.

Quick reference