Tidio

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me the shop chat with an assistant that replaces Tidio.

The useful half is answering the same five questions forever — where is my order, what is the return policy, do you ship here. Build for those and hand everything else to a person quickly.

STACK
- Node 20+ with Fastify, server-rendered HTML
- SQLite through better-sqlite3, WAL mode, FTS5 for the answer base
- A local model through Ollama, or your own key, for the assistant
- Server-sent events one way, POST the other
- Caddy in front

THE DATA MODEL
- conversations, messages, agents as in any chat
- answers: id, question, body, tags, updated_at — your own written answers, not scraped
- bot_turns: id, conversation_id, question, retrieved_ids_json, answer, confidence, handed_off, at
- orders_view: a read-only lookup into your shop by order number and email

THE ASSISTANT, BUILT THE ONLY SAFE WAY
- **Retrieval, not generation from memory.** Search your own answers, and let the model rewrite the retrieved text into a reply. It must never invent a policy
- Every reply cites which answer it used, visible to the agent and available to the visitor
- **If nothing retrieves above a threshold, hand off. Do not guess.** A confidently wrong answer about a refund policy is a chargeback and a bad review
- Never let it state a price, a delivery date or a policy that is not in the retrieved text

ORDER STATUS, THE MOST ASKED QUESTION
- Look up by order number plus the email on the order. **Never by order number alone** — sequential order numbers mean anybody can read anybody's order
- Return only what the customer already knows plus the status. No addresses, no payment details
- Rate limit the lookup hard, per visitor and per order number

HANDING OFF WELL
- A visible, always-available 'talk to a person'. Never make somebody argue with a bot to reach one
- Show honest availability; outside hours take an email and say when they will hear back
- The agent sees the whole bot conversation and what it retrieved, so they do not start from nothing

THE WIDGET
- Under 20KB, shadow root, loaded after the page, keyboard accessible, announced to screen readers
- Opens by itself at most once per visit and remembers a dismissal

MEASURING WHETHER THE BOT HELPS
- Deflection is the wrong metric on its own. Measure **handed-off rate, and the rating of conversations the bot handled versus a person**. If the bot's ratings are worse, it is costing you customers to save agent minutes
- Log every question with no good retrieval; that list is exactly what to write next

WHAT MATTERS MOST
Retrieval-only answers with a hand-off threshold, and order lookup that needs the email. One stops the bot lying; the other stops it leaking.