Create the Assistant people will actually meet

The Base identity

An Assistant is the operating profile behind ordinary conversation.

It combines an administrator label, visitor-facing identity, first message, behavioral instructions, and an OpenAI model. Create one profile you can explain before creating alternatives.

What an Assistant is in SmartSite

SmartSite stores Assistant profiles locally in the WordPress option ai_chat_assistants. Each receives an ID beginning with asst_, but that naming style does not make it an object in the OpenAI Assistants API. The plugin uses the selected local profile to construct requests through OpenAI’s Responses API.

This distinction prevents a great deal of confusion. You manage the profile in WordPress, not in an OpenAI Assistants dashboard. Editing it changes local configuration. Knowledge remains in the site’s shared vector store rather than being copied into each profile. Tools may be assigned to all profiles or one profile ID.

Creating a profile makes it the active Assistant immediately. If a production site already has a working profile, creating a candidate is therefore a live activation event—not a harmless draft action. Prepare and test in Local/staging or during a controlled window.

Every field on the create form

The five decisions that make one Assistant profile
Field	Where it appears	Write or choose it deliberately
Internal Name	Assistant cards and administration only.	Name the purpose and environment: “Website Support – Production.” Do not put a public greeting here.
Display Name	Shown to visitors in the chat widget header.	Use the identity people should trust and recognize, such as “Acme Website Assistant.” Avoid pretending it is a named human.
AI Model	Used for Responses API generation.	Select a text model available to your OpenAI project and verified with this runtime. The complete API model list may contain unsuitable entries.
Welcome Message	First Assistant message when the conversation opens.	Set expectation and invite a useful question. Keep it short; it is not the full instruction set.
System Instructions	Private behavior guidance included in runtime requests.	Define role, scope, source use, uncertainty, tone, escalation, and Tool boundaries. Never store secrets or rely on prose for security.

On the full create form, the model list is fetched from OpenAI and the rendered default is gpt-4o-mini. In the Setup Wizard, the curated list defaults to gpt-4.1-mini. Defaults are starting choices, not a promise that one is best for your language, latency, quality, or budget. Compare with a fixed question set.

Name the public identity honestly

The Display Name affects trust before any answer is evaluated. A label such as “Marko” can imply a human is present. “SmartSite Assistant” or “Acme Support Assistant” signals automation more accurately. Pair the name with a Welcome Message that says what it can help with and does not promise instant human service.

Good:

Hello! I’m Acme’s website assistant. I can help with our services and public policies. What would you like to know?

Risky:

Hi, I know everything about your account and can solve any problem instantly.

The second message claims private access and universal capability before the system has earned either.

Create a profile that is ready to test

Confirm the regular OpenAI key. The Assistants screen loads the model list through that project credential.
Write the scope on paper. Name the public audience, supported questions, source of facts, and human next step.
Open Assistants → Create New Assistant. Remember that Save activates the new profile.
Choose distinct names. Make Internal Name operational and Display Name visitor-friendly.
Choose a tested model. Do not assume every model returned by the API supports the plugin request.
Write a truthful greeting. One or two sentences should be enough.
Write bounded instructions. Tell it what to do when knowledge is missing and which actions require explicit visitor intent.
Create the Assistant. Confirm its card immediately shows Active.
Restrict frontend access. Use a test role/user before exposing the new identity.
Start a fresh conversation. Test direct, ambiguous, out-of-scope, and missing-information questions, then inspect Chat History.

A practical first profile

Internal Name: Website Support – Candidate

Display Name: Acme Website Assistant

Welcome Message: Hello! I can help you find information about Acme services and policies. What can I help with?

Instructions should then explain that it answers public service and policy questions, uses synchronized knowledge for Acme-specific facts, avoids inventing prices or eligibility, asks one useful clarification when needed, and directs account-specific cases to the approved contact route.

Notice what does not belong here: API credentials, hidden discounts, staff-only escalation numbers, webhook authentication, a copy of every policy, or instructions to bypass security. Knowledge and Tools have their own controlled systems.

How the profile combines with the rest of SmartSite

An Assistant does not work alone
Neighboring system	Relationship	Common misunderstanding
Knowledge Base	The active profile can search the shared synchronized library when runtime permits.	Creating another Assistant does not create a separate library.
Tools	Enabled Tools assigned to All Assistants or this profile can be offered to the runtime.	Assignment limits profiles, not visitor identity or authorization.
Agent Flows	The active profile becomes Base Assistant; published specialists add scoped routing.	Turning on Advanced Agents does not make the local Assistant obsolete.
Settings/Design	Other screens control landing copy, avatar, access, visual style, and position.	Display Name/Welcome Message do not own every word or visual element in the widget.
Analytics	New conversations record the model, profile identity, response, retrieval, Tool activity, and errors where available.	A saved profile is not verified until a real trace uses it.

The model selector requires judgment

The full Assistant screen lists models returned by OpenAI. Availability in that list does not guarantee compatibility with the Responses API request, desired tools, or your OpenAI project permissions. If a chosen model fails, select a known compatible text model rather than changing knowledge and instructions at the same time.

Evaluate quality, latency, and token/cost evidence with the same representative questions. Model choice can affect instruction following and language quality, but it cannot repair missing facts or unsafe Tool design.

What success looks like

The card shows the intended profile Active. A fresh visitor sees the intended Display Name and Welcome Message. Direct questions follow the written scope, site-specific answers retrieve approved knowledge, out-of-scope questions remain honest, and no action Tool runs from hypothetical language. Chat History identifies the expected model/profile without error.

Do not test only “Hello.” Greetings prove little about grounding, boundaries, or action safety.

Give the profile an owner

Record who approves its public identity, who owns its Instructions, and who can activate a rollback profile. Review it when services, policies, brand language, escalation routes, Tools, models, or Advanced workflows change. A profile can remain technically valid while becoming operationally misleading.

The owner should also know where the Assistant ends. They do not need to edit webhook code, but they must recognize when a requested behavior depends on Knowledge, a Tool, access control, or a specialist rather than adding another instruction sentence.

See the creation desk

Screenshot capture pending /images/screenshots/assistants/understanding-create.webp

Understand and Create Assistants

Capture: Show a completed Create New Assistant form using fictional content, with the model selector closed and no secrets visible.
Show: Name, Display Name, Welcome Message, Instructions, Model, Save 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