Providing context to the Chat API before a conversation

Adding a bit more to Paul’s answer, and using your example.
Use the System role for context maintenance. Itemize, numbered is better, concise instructions, mind punctuation as separators:

System:
Please follow the six instructions below:
1. Perform assistance for a US second-grade math learning;
2. Provide a step-by-step process for a solution to any math problem presented in the User role;
3. Do NOT provide any final or partial answer to a math problem;
4. Use US second-grade English language level;
5. Provide Spanish translation in parenthesis of all mathematical terms presented on US second-grade or higher, in problem or solution;
6. Check ALL final or partial answers to the problem no matter the solution process used in the User role.

Some GPT-3.5 versions have a weakened System role importance. In order to make a model-free conversation - it is suggested to do this in the User role:

User:
Follow the instructions in the System role always.
Keep those instructions in context all the time.
...
"second-grade math problem text..."

“I” and “you” work nicely in the User prompt - it is usually expected by the models in a more natural language. In order to keep the conciseness of the instructions in the System role, the pronouns are unnecessary, more robotic command sentences shall be enough.

The models give low importance to the Assistant role concerning to context maintenance.

The User role shall have more detailed information in order to solve the math problem. If you notice that the model “forgets” the context-maintenance instructions, you can repeat the sentence Follow the instructions in the System role... in the User role header to reinforce the System authority over the context of the conversation.

I hope this helps. Please, let us know the results - there are more threads in this forum with these tactics with relative success.

5 Likes