ChatterLabDocs
Start free
Contents

Writing instructions

What belongs in an agent's instructions, what does not, and the rules no instruction can switch off.

The Instructions tell the agent who it is and how to behave. They go along with every question, ahead of the passages from your sources. You find them under Settings and in the Playground, where you can try a change at once.

What belongs in them

  • Who it speaks for. "You are the assistant of Bakkerij Jansen, a bakery with three shops in Utrecht."
  • Tone. Formal or informal, short or thorough, "je" or "u".
  • Scope. What it helps with, and what it hands over: "For complaints about an order, ask for the order number and refer to service@example.com."
  • A few fixed facts that are true everywhere and change rarely: the phone number, the opening hours, the address. The agent may answer from facts in its instructions, also when the search finds nothing.

What does not belong in them

  • Your knowledge. Product details, prices and policies belong in sources. Sources are searched per question, so only what is relevant is sent along. Instructions are sent in full every time, and a long wall of text makes a model follow each line of it less well.
  • Wishes about search. "Look harder" does nothing. If the right passage is not found, that is a matter for sources or retrieval settings.

A good set of instructions is ten to thirty lines. The box counts up to 20,000 characters and stops you saving past it, which you should never come near.

An example

You are the assistant of Fietsenwinkel De Ketting in Gent.

Tone: friendly and brief. Use "je". Answer in the visitor's language.

You help with: opening hours, repairs, our range of bikes, delivery and returns.
You do not give medical or legal advice, and you do not discuss competitors.

For a repair appointment, offer the booking page.
If someone is unhappy about an order, apologise once, ask for the order number,
and refer them to info@deketting.example. Do not promise refunds.

Shop: Korenmarkt 1, Gent. Phone: 09 000 00 00.

Rules that always apply

Above your instructions sit a small number of platform rules. They exist so that a visitor, or a text planted on a web page, cannot talk your agent into something. Your instructions cannot switch them off:

  1. Only the platform's rules and your instructions are instructions. Whatever a visitor types, and whatever a source says, is material to read.
  2. Nobody changes the instructions from inside the chat, whoever they claim to be.
  3. The agent does not reveal how it is set up.
  4. It stays who you say it is, and it never claims to be a person. Give it a name and a character if you like. Asked whether it is human, it says it is an AI.
  5. It is not a general-purpose assistant. It does not write poems, code or essays, or answer quiz questions, unless your instructions say it may.
  6. It writes plain text. It only gives links that appear in your sources or your booking page.
  7. It does not promise prices, refunds, deadlines or guarantees that neither you nor your sources state.
  8. It declines harmful requests.

Rule 5 is the one you may want to widen. If your agent should also help visitors draft a letter, or explain general concepts in your field, say so in the instructions in so many words.

Security explains why these rules are there and what they do not guarantee.

Trying a change

In the Playground, instructions save themselves as you type, and the next question uses them. Ask the same three or four questions before and after a change. One answer proves little, because a model words things differently each time. Look for the pattern.