What are Prompts?

Like you would expect prompts in Turingpaper are natural language instructions of what an agent should do. They can contain "parameters" or "placeholders" put between double curly-brackets, and they always have a name.

Here is an example prompt called search_result_summary with one parameter called search_topic:

Use the tool `internet_search` to find relevant material for the search
topic `{{search_topic}}` and write a summary of all the search hits.

Prompts are especially useful when they contain parameters because you or the agent can give different values to the parameters and the prompt will be evaluated by an agent with the values you pass filled in.

For example: Imagine that you use the search term "Top SEO considerations with LLMs" as the parameter search_term in the prompt defined above, search_result_summary.

The full prompt text that the agent would see and carry out as instructions would be:

Use the tool `internet_search` to find relevant material for the search
topic `Top SEO considerations with LLMs` and write a summary of all 
the search hits.

Prompts Can Use Other Prompts

Turingpaper allows a prompt to call other prompts as if the prompts were tools. This is important because you can easily break down complex instructions into multiple simple instructions that refer to each other.

A common pattern in agentic AI automation is to use one prompt as the "manager" and another prompt as the specific task to perform. This is used often when you need to perform the same specific task on multiple inputs, like a folder full or reports.

Use the tool `file_list` to find all the reports in the folder
`{{report_folder}}`, and for each one call `process_one_report`.

Read the report in `{{report_filename}}` and find the conclusion,
then double check if each claim made in the conclusion has supporting
evidence in the body of the report.

When to Use Prompts?

You should consider using a prompt before you consider creating a new custom Agent or before writing a Function.