Your first trustworthy conversation

The shortest responsible path

Build one good conversation before building every feature.

The first milestone is not a public widget. It is a permitted test visitor asking a known question, receiving an accurate answer from the intended Assistant and source, and leaving a readable trace in Analytics.

The first good conversation—not the first public launch

Quick Start deliberately postpones Tools, WhatsApp, Advanced Agents, elaborate design, broad visibility, and a large document library. Those features can be valuable, but each adds a new place for setup, security, and ownership to fail.

Your baseline uses one regular OpenAI API key, one active Assistant, two or three authoritative pages, restricted access, and a fresh frontend conversation. Once this path works, every later change has something stable to compare against.

Build the smallest complete Assistant

Open Setup Wizard. The Welcome step explains the required and recommended stages.
Save the regular OpenAI API key. Use a scoped key for the intended project. The wizard saves it but does not validate it live.
Create one Assistant. Give it an internal name, public Display Name, short Welcome Message, clear Instructions, and a model intended for the Responses API.
Select two or three pages. Choose durable sources whose answers you know. Select Save Selection & Sync and inspect the result.
Skip Tools. Presets install enabled; leave them for the dedicated review unless you can disable and test them immediately.
Finish Setup. This records wizard completion and returns to the main plugin area.
Restrict widget access. In Settings, allow only a test role or user before opening the frontend.
Ask a controlled question. Use a separate/private browser and wording whose answer exists clearly in one synchronized source.
Read Chat History. Confirm the Assistant, model, result, knowledge activity, duration, and errors correspond with the test.
Configure Security and privacy. Only then consider widening access.

Creating an Assistant makes it active immediately. Saving selected pages is not enough: synchronization makes page knowledge available remotely. If a selected page is Pending or Not synced, stop and resolve that before judging answer quality.

A first instruction set that is useful

Do not begin with a personality novel. Start with a short operating brief:

You are the website assistant for Example Company.
Help visitors understand our public services and policies.
Use the synchronized website knowledge for company-specific facts.
If approved information is missing or unclear, say so honestly.
Do not invent prices, availability, eligibility, or commitments.
For account-specific help, direct the visitor to the contact page.
Reply clearly and in the visitor's language.

Replace the company, scope, and escalation path. This creates observable behavior without pretending instructions are access control. Keep secrets and internal procedures out of the field.

Choose first sources that can pass a test

A small launch library
Source candidate	Why it is useful	Why to postpone it
Service overview	Provides a clear explanation and visitor vocabulary.	Postpone if it is a thin campaign page or makes outdated claims.
Contact/location page	Easy to verify with direct questions.	Postpone if it contains several conflicting locations or blank dynamic fields.
One important policy	Tests conditions, dates, and honest uncertainty.	Postpone if legal/content ownership has not approved the current wording.
Large handbook or archive	May eventually add depth.	Too broad for a baseline; duplicates and old facts make diagnosis difficult.

Ask one direct question and one synonym: “When can I cancel?” and “What is the rescheduling deadline?” Then ask something the source does not cover. A good Assistant should handle both knowledge and limitation.

The five-question launch gate

Do not widen visibility until every answer is yes
Gate	Evidence	If no
Can the intended visitor open the widget?	Restricted frontend test in a separate browser.	Fix Settings → Access, shortcode/page placement, or frontend assets.
Is the intended Assistant active?	Active badge plus its Display Name and behavior in a new conversation.	Activate the correct profile and restart the test context.
Is approved knowledge available?	Sources are Current/completed and the intended source appears in diagnostics.	Synchronize, wait for indexing, or improve the source.
Does the answer remain inside scope?	Direct, synonym, missing-information, and out-of-scope questions behave honestly.	Rewrite instructions/source and repeat one change at a time.
Can the team see and stop problems?	Chat History works; an admin knows visibility, Assistant, Tool, and runtime kill switches.	Finish operations and incident ownership before launch.

What to leave for day two

Add a Tool only when you can state what it returns or changes, who owns the destination, how it validates input, and how it is disabled. Install presets one at a time and disable immediately for review because they arrive enabled. Add WhatsApp only after the web experience is stable; otherwise channel issues obscure Assistant issues.

Use Advanced Agents when one general Assistant has proven insufficient for distinct specialist tasks. The Base Assistant and shared knowledge remain relevant. Build a small flow, test routing without action execution, publish a revision, then observe traces.

Design can be refined after the conversational baseline. Welcome and display identity come from the Assistant; landing text, avatar, colors, position, and access live elsewhere. Changing the wrong label is a common source of wasted work.

How to react when the first test fails

No widget: check visibility/access and frontend placement before touching OpenAI.
No response: confirm the regular API key, active Assistant, chosen model, and exact error.
Generic answer: confirm knowledge search ran and the page is Current.
Old behavior: open a new private conversation; previous response context can persist.
Wrong tone but right facts: revise Assistant Instructions, not the source.
Correct in admin, wrong in public: ensure the frontend test uses the same site, active Assistant, runtime, and access configuration.

Do not change five settings at once. Preserve the question, change one owner layer, and compare the next trace.

A production launch is a separate decision

Before selecting Everyone, review privacy disclosures, logging and retention, message length/rate controls, blocked/detection behavior, accessibility, mobile layout, support ownership, OpenAI cost expectations, and any enabled Tools. Run the regression questions again after the production access change.

“Finish Setup” means onboarding state is complete. It does not certify external API validity, synchronized content accuracy, mail delivery, webhook authorization, or safe public visibility.

The final wizard checkpoint

Screenshot capture pending /images/screenshots/start/quick-start.webp

Quick Start

Capture: Show the completed wizard checklist immediately before selecting Finish Setup, using sanitized assistant and page names.
Show: Wizard progress, completed steps, Finish Setup button
Viewport: Desktop, 1440 × 900
Annotate: Use numbered callouts only for controls referenced in the procedure.
Redact: OpenAI keys, tokens, secrets, personal information, private URLs, IP addresses, and conversation text

Continue deliberately

Need every wizard detail? Walk through all six stages.
Need better behavior? Write the Assistant’s operating brief.
Need stronger sources? Understand the Knowledge Base.
Preparing public access? Review privacy and production.