Helpful Tips

Some tips to be productive building your agentic AI automations in Turingpaper:

1. Start with the Generalist agent then create your own

You can create custom agents in Turingpaper with your own specific instructions, but Turingpaper comes with one default agent called the "Generalist".

The Generalist agent has pre-built instructions of how to use many (but not all) of the tools available in Turingpaper, and is a good agent to start with when creating new automations.

Most of the context for the automation should be in your prompts rather than in the agent instructions. Only instructions that you find yourself repeating over and over again can be moved into the agent's instructions.

You can clone the Generalist and customize the instructions to fit your specific automation, and then save it with a new and memorable agent name like "Sales Analyst". After customizing and saving, you can use your Sales Analyst agent in your project.

2. Make tools accessible to the AI by mentioning the tool names

An agent in Turingpaper is allowed to use only the tools specifically mentioned in the agent's instructions, prompt instructions, or a chat message sent by you (the human). The tool name must appear between backticks, for example:

Use the tool `postgres_exec` to query the number of users that have a
credit balance of over $100 in the table user_credit_balance.

3. "Prompt Program" using English

At the heart of Turingpaper automations are prompts. Prompts in Turingpaper are more than just instructions!

Prompts are automatically usable as tools, which allows prompts to call other prompts, which allows complex instructions to be broken down into multiple smaller instructions, for example, look at the following two prompts. The first one uses the second one as a tool:

For each of the following companies use `research_company_value_prop` 
to research the company, and then consolidate all the reports into a 
table comparing the results.
 
List of companies:
 - https://company-a.com
 - https://company-b.com
 - https://company-c.com

Use the tool `fetch_links` to fetch the links from `{{company_url}}` and 
for each link that is related to their product offering use
`fetch_content` to get the content. Read the content of the product
related pages and create a report about the company's value proposition.

4. Use functions when processing lots of data

LLMs like gpt-4.1, o3, and o4-mini are good at following instructions, but they have their limits.

Turingpaper automatically converts both prompts and Turingpaper JavaScript functions into tools that can call each other and be used by AI agents.

When an automation needs to process a lot of data in the same way consider creating a Turingpaper JavaScript function that will perform the repetitive parts of the automation.

Use the tool `file_list` to list all the folders in `/job_openings`.
For each folder use the tool `evaluate_resumes`.

function evaluateResumes(folderName) {
    const fileNames = list_files(folderName);
    for (const resumeFilename of fileNames) {
        evaluate_one_resume(folderName, resumeFilename);
    }
}

First, read the job description in `/{{folder_name}}/job_description.md`.
 
Then read the resume in `{{resume_filename}}`.
 
Then evaluate the job fit based on the experience described
in the resume and save your analysis by creating an analysis
file in `/{{folder_name}}/candidates/analysis_{{resume_filename}}`.