Build and find datasets
Build a dataset from synced tables, find the approved dataset that answers a question, and read a table.
Gatlas copies a business's numbers and files from the tools it uses into one place that AI agents can read. Its synced tables hold the raw rows; a dataset turns them into a table that answers one question, and agents read the approved ones first. This playbook builds datasets, finds the one that answers a question, and explains a table.
How to run
Run the section the person's prompt names. If it names none, ask what they want to do and pick the section that fits. The prompt may name a company or a table; use them, and ask for what it leaves out.
Do each step yourself with Gatlas's tools. Pause only for an approval the person must give. When the person only asks questions, or their key is read-only, use Find a dataset or Find a table, never Build a dataset.
Ask one question at a time, and wait for the answer before the next. Never ask for a key or a password in this chat. Before you change anything in Gatlas, tell the person exactly what will change and wait for their yes. Never delete anything.
Before you start
Check that you can see Gatlas tools such as list_companies. If you cannot, stop and walk the person through Connect Gatlas to your agent first, then come back here.
Build a dataset
Help the person build a dataset in Gatlas, for the company the prompt names when it names one. A dataset turns synced data into a table that answers one question, such as monthly revenue by company.
Ask which question to answer first. Use list_tables and describe_table to find the tables that hold it, test the SQL with query_sql, and show the person the SQL, a few sample rows and the AI context you would write. Name the dataset after what it holds and set it to refresh daily.
Once they say yes, create it with create_dataset.
Before you change anything in Gatlas, tell the person exactly what will change and wait for their yes. Never delete anything.
Find a dataset
Help the person find the approved dataset in Gatlas that answers a question. Ask what they want to know. Use search_datasets to find likely datasets and describe_dataset to read what one row is and what each number means.
Then answer from it with query_dataset, and tell them which dataset and which columns the answer came from. Do not change anything in Gatlas.
Find a table
Help the person find the synced tables in Gatlas that answer a question. Ask what they want to know. Use list_tables to see each company's tables and describe_table to read the likely ones.
Then tell them which table holds the answer and which of its columns matter. Use list_datasets to check whether an approved dataset already answers it, since agents should read that first.
Explain a table
Explain the Gatlas table the prompt names, or the one the person has open. Use describe_table to read its columns, row count and dates, and tell them in plain words what one row is, which columns matter and what to watch out for.
Then use list_datasets to tell them whether an approved dataset already answers questions from it.