Tracked prompts

A tracked prompt is one question, asked on a schedule, on every assistant, with the answer stored in full. It is the unit of measurement here, and the quality of your prompt list sets the ceiling on everything else.

A prompt can also be marked as a control, tracked and never worked, or run once as a research prompt rather than on a schedule.

What makes a prompt work

An answer only contains brand names when the question is specific enough to need them. The qualifier does that work: the team size, the industry, the budget, the country, the constraint.

"Project management software" gets a lecture. "Project management software for a 10 person construction firm that bills hourly" gets five brand names and a reason for each.

What to track

Cover the journey rather than the category.

• Early questions, where the buyer is describing a problem and not a product.
• Choice questions, where they are picking between options.
• Comparison questions, naming two brands.
• Brand questions, naming you, and naming each serious rival.

The brand questions are the ones teams skip and the ones that tell you how assistants characterise you when they already know who you are.

Editing and retiring

Editing a prompt's wording changes what is being measured, so the history before and after is not strictly comparable. For a small rewording that is fine. For a real change of question, add a new prompt and retire the old one so each trend line means one thing throughout.

Retiring stops the runs and keeps the history.

Unlimited, and still worth pruning

There is no cap on tracked prompts on any plan, because the model calls are on your key. That makes the real limit your attention and your provider bill: twenty prompts you read beat two hundred you do not.

For a one-off study of two hundred questions, use research prompts instead: they never occupy a tracked slot, and the whole batch is costed before it runs.