Summation Start

Walk a new user from zero to a first useful answer in plain language. No HTML artifacts or welcome steppers. Onboarding is MCP: the signin skill (host OAuth) is the product sign-in, and discovery below uses MCP tools. CLI-first applies after sign-in whenever a shell is available — data work then prefers sumcli (≥ 0.1.3, install at first need after an explicit yes — ../api/references/sumcli.md), which needs its own one-time sumcli auth login; MCP stays the fallback for shell-less hosts. For “what’s supported?”, always fetch https://docs.summation.com/llms.txt (see ../api/references/product.md).

Checklist (optional one-liner in chat)

1 Connect → 2 Discover → 3 Meet the analyst → 4 First report

Step 1 — Connect (sign in)

Run the signin skill. Show who they are (name/email, org) in ordinary words.

Step 2 — Discover (hard gate)

MCP first (do not stop early): list_data_connections, list_connection_datasets, list_projects, search_tables as needed. Treat imported project tables and attached connection datasets as analyzable data.

Gate open when either holds:

• at least one connection with attached datasets (friendly names — never leave auto *_dataset_N), or
• at least one imported / project table with a clear analyzable name

Then give a short source map and continue.

Gate closed only when both are missing — no attached datasets and no imported tables. Then stop and offer:

1. Connect a database/warehouse → the connect skill (confirm sources via llms.txt / product docs)
2. Import files into a project → file import tools, then re-list tables
3. They already connected or imported in the web app → re-list and continue

Connection exists but nothing attached and no imports → browse and attach with friendly names (table names), or import. Do not claim readiness until the gate opens.

Step 3 — Meet the analyst

1. Ensure a project (list_projects / get_default_project / create with consent).
2. Project catalog: attach agreed business tables if empty.
3. ask_analyst: introduce what you see + 3 concrete report ideas from real table names. Tell the user the analyst is working.

Step 3b — Opportunities from recent chats (optional, once)

After the gate is open or after a clear “no data yet” path, offer once:

I can also look at recent local chats in this IDE for workflows Summation can take over (scan stays on your machine). Want that?

• Yes → run the opportunities skill (consent + local scan + catalog-grounded list). Prefer their pick for step 4.
• No → one line: anytime the opportunities skill — then continue.
• Do not scan without consent. Do not dump transcripts.

Step 4 — First report

Numbered ideas (from catalog and/or opportunities). On yes → report skill (markdown) → offer validate. Next steps in plain language.

Voice

• Outcomes only: no endpoints, OpenAPI, or raw tool-id dumps.
• Never claim “added” / “imported” without a final list/check.
• No HTML welcome visuals.

Rules

• Analyzable data = named datasets / tables, not system tables alone and not “browsable but unattached.”
• Passwords never in chat; connect skill owns secrets.
• Keep momentum: short steps, clear waits on long tools.