Writing Chatbot Personas and Guardrails That Don’t Embarrass You

After your knowledge base is filled, two text fields determine whether the bot feels helpful or annoying: persona and guardrail. Both are short, easy to overlook, and their effect is visible in every conversation.

Persona: who is speaking

The persona explains the bot’s role, its tone, and its limits. Often, this field is filled with adjectives that can’t be acted upon. Writing friendly, professional, and informative doesn’t change anything, because those three words apply to almost all answers.

What works are instructions that can be checked for correctness.

Kurang berguna:
Kamu adalah asisten yang ramah, profesional, dan informatif.

Lebih berguna:
Kamu asisten pra-penjualan untuk Kopi Rakyat, pemanggang kopi di Bandung.
Pakai bahasa Indonesia santai tapi sopan, sapa dengan "kamu".
Jawab maksimal 4 kalimat, lalu tawarkan satu langkah lanjutan.
Kamu tidak berwenang memberi diskon, mengubah pesanan, atau menjanjikan
tanggal kirim. Untuk hal itu arahkan ke tim lewat WhatsApp.

The difference lies in three things: the business context is mentioned, the length of the answer is limited, and what is not allowed is written explicitly. That limit of authority is the part most often missing, yet it’s what prevents the bot from promising things you can’t deliver.

Additional context: facts not on the site

This field is where you put things that are frequently asked but never written on any page. Operating hours, service areas, the fastest way to contact the team, temporary policies.

Its content is always sent to the model, so restrain yourself. Piling up your entire brochure here makes important parts sink and the cost per conversation increases. If information deserves to be read by many people, its place is on your site pages, not in this field.

Guardrail: the sentence when the bot doesn’t know

Guardrails are used when the search doesn’t find enough material. This is the most vulnerable moment in a conversation, because this is where a bot without a guardrail starts to make things up.

A good guardrail acknowledges it doesn’t know, then provides a concrete way out. A bad guardrail stops at the first sentence and leaves the visitor at a dead end.

I don’t have that information on our pages. For questions about stock and shipping, our team can be contacted on WhatsApp 0812 xxxx xxxx from 09.00 to 17.00.

• Mention a follow-up path that people actually monitor.
• Use the same pronouns and tone as the persona, so it doesn’t sound like a system error message.
• Avoid blaming the questioner. Sentences like your question is unclear make people leave.
• Don’t promise the bot will call back. The bot can’t do that.

Testing Before Publishing

The preview widget on the Widget screen (Editor tab) is for this stage. Before pasting the script onto your production site, run at least four types of questions:

1. 1Questions you know the answer to are on the site. Check the content is accurate, not just sounding right.
2. 2Questions where the answer clearly isn’t there. Check the guardrail really appears.
3. 3Questions completely off-topic, for example about the weather. Check the bot politely refuses.
4. 4Questions that solicit promises, like asking for a discount. Check that authority limits are respected.

Maintaining it after launch

Chat history is the most honest source of improvement. Read the first week of conversations and look for two patterns. First, questions answered by the guardrail repeatedly, meaning content needs to be written. Second, correct answers with the wrong tone, meaning the persona needs adjustment.

Those two patterns are usually enough to make the bot feel much more mature in the second month than on the first day, without changing anything on the technical side.