Help · Guides · OperatorChat

OperatorChat — design a chat persona that actually sounds like you.

App: /apps/operatorchat/ Cost: 3 credits per persona Time: 7 min · ships .md persona export

Most chatbots fail on tone. They're either too corporate (kills warmth) or too friendly (kills authority). OperatorChat ships a complete persona pack — system prompt + voice notes + 5 sample exchanges — so you can drop it into any chat platform and it sounds intentional from message one.

Open OperatorChat

Go to /apps/operatorchat/. Sign in at /portal/ if needed.

3 credits per persona. Plus's 75/month = 25 personas. For most brands you'll design 1–2 personas total (one customer-facing, maybe one internal) so this is a low-volume, high-stakes tool.

Brief the persona: brand + role + scope

The form takes:

Brand — what's the company / product? One line.
Persona role — sales bot, support agent, onboarding guide, internal HR helper, etc.
Scope — what's it allowed to answer, and what should it escalate? Be explicit. "Answers shipping + product questions; escalates returns + refunds to a human" beats "is a helpful chatbot."
Voice cues — adjectives + examples. "Direct, slightly cheeky, never uses 'kindly' or 'reach out'" is the kind of input that produces a sharp persona.

Tip — write the "never says" list. The cheapest way to make a chat persona feel intentional is to name what it WON'T say. "Never says 'absolutely,' 'I'd love to help,' or 'just to clarify'" instantly removes 80% of generic-chatbot energy.

Read the system prompt + voice notes block

You get back:

System prompt — ready to paste into Claude, GPT, or any chat-completion API's system field.
Voice notes — the human-readable summary of how the persona should sound. Useful for explaining to non-eng teammates or future-you.

The system prompt is structured: role definition → tone rules → "never says" list → scope boundaries → escalation triggers → format constraints.

Inspect the 5 sample exchanges

OperatorChat doesn't just give you the prompt — it shows you 5 example user-message-and-persona-reply pairs covering different conversation types:

A normal in-scope question
An edge case that requires soft escalation
A confrontational / frustrated user message
A vague / context-light question
An out-of-scope ask that requires polite redirect

Read all 5 before deploying. If any reply feels off, refine your brief and regenerate. Cheaper to iterate on the persona than to ship a bad one and lose customer trust.

Push the prompt further in PromptGen (optional)

If your chat platform has constraints (token limits, output-format requirements, function-calling), hit the Refine in PromptGen button. It pre-loads the persona system prompt + lets you add platform-specific tuning.

OperatorChat vs PromptGen	When to use which
OperatorChat	Designing the persona itself — tone, scope, behavior. Optimized for ongoing conversations.
PromptGen	One-shot tasks, transforms, classifiers. Optimized for "input → output" not "back-and-forth."

Export and deploy

Hit Export as Markdown for the full persona pack as one .md file. Drop it into:

Claude Projects custom instructions
GPT Custom GPT setup screen
Intercom Resolution Bot, Drift, Tidio, etc. — paste the system prompt into their AI config
Your own AI chat integration via Claude API / OpenAI Chat Completions API

When OperatorChat is the wrong tool

You need a knowledge-base chatbot — OperatorChat designs the persona, not the knowledge retrieval. You'll still need a RAG / vector DB layer if the bot answers from documentation.
You need IVR / phone bot scripts — voice-first design is different. Use OperatorChat for the persona then adapt for voice (shorter turns, no markdown, no emoji).
You need a basic FAQ widget — OperatorChat is overkill. Use plain text + simple if/then.

What's next? Most brands deploy one persona to production and iterate it weekly for 2–3 weeks based on real conversation logs. Save your initial .md export, then save each iteration as operatorchat-v2.md, v3.md — so you can A/B versions and trace why responses changed. After ~30 days of real chat data, the persona is calibrated and you can leave it.