TopolearnSeparating your rules from the user's text is the basis of predictable behaviour.
Most apps have two kinds of text: your fixed rules ("You are a support assistant for Acme; answer briefly") and the user's changing question. The Responses API gives each its own place.
instructions holds the rules. It works like a system prompt: the model treats it as guidance from you, the developer, and gives it priority over user text.
input can be a list of messages, each a dict with a role and content. Roles you will use:
user: what the person typed.assistant: an earlier reply from the model.system: developer guidance placed inside the input list, an alternative toinstructions.
A list lets you include earlier turns, or a worked example of the answer style you want.
A gotcha: instructions apply only to the request they are sent with. When you chain requests with previous_response_id (a later lesson), earlier instructions do not carry over, so send them every time.
Write instructions like a brief to a new colleague: who the model is, what the task is, what format to use, and what to avoid. Explain why a rule exists; models follow reasons better than bare commands.
Example
from openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-6-astra",
instructions=("You are a friendly cooking assistant. Answer in at most "
"three sentences, because users read on a phone."),
input=[
{"role": "user", "content": "How do I keep rice from sticking?"},
],
)
print(response.output_text)Remember
instructions= your rules;input= the conversation- Input messages use roles: user, assistant, system
- Resend instructions on every request; they don't carry over
Watch out for
- Assuming instructions from the first request still apply when chaining with previous_response_id.