Skip to main content
You can automate a column by writing a prompt, a set of instructions for an AI agent to follow and eventually produce a result to write into a cell. The prompt is ran independently for each row in the column, using other values in the row as inputs. Because each prompt may be ran thousands of times, we’ve produced a specialized prompting language to help you write prompts that are robust and behave consistently at scale.

​
Referencing other columns

A prompt is only useful if it depends on other columns; otherwise, it’d output the same thing in every row. You can reference columns using the syntax Column Name. For example:
Calculate the compound annual growth rate (CAGR) for the company’s revenue using Starting Revenue, Ending Revenue, and Number of Years. Express the result as a percentage rounded to two decimal places.
When you run a group of columns at once, each column will wait until all its dependent columns have finished running.

​
Using tools

If you don’t reference any tools, the agent can only do the kind of things you can do with Excel formulas. To access the internet or run AI analysis on each row, you need to instruct the agent to use our special tools. This is done using the syntax /ToolName. The simplest tool is /AskLLM, which just passes a prompt to an LLM and gives back the result. For example:
/AskLLM to analyze the sentiment of Transcript and classify it as “Positive”, “Neutral”, or “Negative”.
You can specify the model you want to use; popular options include Sonnet 3.5 (default), GPT-4o, and Perplexity. Perplexity is a model with access to the internet, so here’s an example that takes advantage of that:
/AskLLM Perplexity “What is the current market share percentage of Company Name in the Product Category market in Country?”
If you want to perform more thorough web research, or have more control over the process, you can use this common pattern:
  1. /Google “average temperature in City name”
  2. /Visit the top 3 results
  3. /Extract the temperature in Celsius. If it’s in Fahrenheit, convert it to Celsius.

​
Control flow

By interweaving tool references within sentences, you can naturally perform conditional logic and even loops. For example:
Grab all the twitter usernames from Blog post content. For each username, /Visit x.com/{username} and /Extract the content of the first three tweets. /AskLLM whether any of the tweets are shitposts (yes or no). If less than half of the users have a shitpost, output “good”, otherwise output “bad”.
Note that with complex logic like this, it’s often helpful to split the prompt into multiple columns.