All guides

AI research columns: how to write prompts that give you usable answers

October 24, 20261 min read

An AI research column runs the same question against every row in a table. Done well, it gives each account a one-line reason to reach out. Done badly, it fills a column with confident filler you cannot send.

Ask one question per column

"Summarise this company" produces a summary nobody reads. "Is this company hiring salespeople right now? Answer yes or no" produces something you can filter on. One clear question per column beats one long prompt.

Pick the output type first

  • Yes/no for filters: "Do they sell to enterprises?"
  • A pick-list for segments: "Which best describes them: agency, SaaS, ecommerce, other?"
  • A number for scoring: "How many open sales roles are listed?"
  • Short text for personalisation: "One sentence on what they launched this quarter."

Typed answers are easier to check. A yes/no that comes back as a paragraph is a sign the question was unclear.

Ground it in a source

Answers from model memory go stale and can be invented. When the row has a website or domain, the research should read that site first. Ask for facts the site can answer, and treat anything else as a guess.

Test small, then run

  1. Run the column on five rows.
  2. Read every answer. Would you send an email based on it?
  3. Tighten the question until four out of five are usable.
  4. Then run the rest of the table.

In Kiln

AI columns use {Column} placeholders in the prompt, return text, number, yes/no, pick-list or JSON answers, and read the row's domain directly when one is present. Each agent comes with five free sample runs for testing prompts, and rows that fail are not billed.

More guides