Tutorials

Four real setups, each built from tools and a prompt, each with a template you can import and change. Start with the one closest to what you need: they also teach the patterns the others build on.

Tutorial	What the visitor gets	What it teaches
Order tracking	"Where is my order?" answered from your shop and the carrier	Asking for a value in the chat, saving it, one tool feeding another ({{data.*}}), picking fields from a big response
Send chats to your CRM	Their enquiry, with a summary, in HubSpot without anyone typing it	Extract tools, an after-chat tool, n8n or Zapier as the glue
Nearest store	Their nearest store in the first answer	A pre-chat form field, a before-chat tool, a tool's result in the prompt
Verified account changes	Change their email or address after an SMS code	Several tools in one task, security your API enforces, answers that are not errors
Logged-in customer support	Billing answers from their own account, without being asked who they are	Signed-in visitors, {{user.id}}, requests people decide on
Product and stock	Live prices, and stock at their nearest store	What goes in the knowledge base and what comes from a tool
Hot leads to Slack	An alert in your sales channel for hot leads only	An after-chat tool with Only when, the AI's labels

Guides: Knowledge base or tool? (where each kind of information belongs) and Testing and monitoring (keeping a setup working as you change it).

Use a template

Every tutorial's setup is a ready-made agent file: the prompt, the tools and any settings they need.

• Dashboard: Settings → Import & export → Use this. It shows what changes, asks for the keys the tools need, then imports.
• Terminal or your coding agent:

helppuff agent templates                      # the list
helppuff agent import order-tracking --dry-run   # what it would change
helppuff agent import order-tracking          # keys come from .env

Then change the addresses to your own APIs, reword the prompt for your business, and test it like a visitor (helppuff ask "Where is my order 1042?"). Or ask your coding agent: "Set up the order tracking template with me".

Patterns that apply everywhere

• Name each tool in the prompt ({{order_lookup}}) at the step where the assistant should use it. A tool used during the chat is offered only when the prompt names it.
• Write the prompt as steps and situations: what to ask, which tool to call, what to say for each result. The assistant follows numbered steps well, and says things in its own words.
• Describe every value the assistant fills in ({{args.order_number}}: "The order number, like 1042 (without #)"). It reads the description to know what to ask the visitor for.
• Answer 200 for expected outcomes. "Not found", "wrong code" or "not verified yet" are answers, not failures: return them as { "found": false } with HTTP 200. A 4xx or 5xx tells the assistant the check failed, and it says it could not check right now.
• Pick what the assistant needs. Tick (or --pick) only the fields it uses: a smaller response means faster, cheaper and more accurate answers.
• Let your API enforce the rules. The assistant follows the prompt, but security must not depend on it: your API decides who may see an order or change an account.