Define What You Are Collecting Before AI Collects It── When the definition comes first, the classification axis is settled and the collected information feeds straight into decisions

01Before building the delivery, the operator's first request was a definition

In September 2026, the operator set out to build a system that would deliver the latest AI research papers every day. The first request to the AI was not for a collection programme or a page layout. It was for a definition of what makes a topic "hot."

The AI began its answer by ruling something out: a single paper attracting attention does not qualify. A hot topic, the AI wrote, is a state where multiple independent signals about the same question rise simultaneously within a short period.

Once this definition was in place, what to measure became clear. The categories used to group those signals were split not by which source provided the data but by who was responding. The design of the collection programme followed naturally from the definition.

02A definition that comes first gives every collected item a built-in criterion

Collecting first and organising later produces a different result from defining first and then collecting. Without a definition, everything goes in, and the work of deciding what matters comes afterwards. With a definition, each item can be classified the moment it arrives.

In the operator's case, the definition contained the condition "multiple independent signals." That single condition ruled out topics that were noisy in only one channel. The need to examine each item one by one after collection disappeared.

03Collecting without a definition multiplies redesign work

When information is collected without a definition of the target, two problems follow. First, every keep-or-discard decision relies on the judgement of whoever happens to be reviewing, with no anchor. Second, if the classification axis needs to change later, the collection itself may have to be re-run.

In the operator's record, the initial attempt split signal categories by data source. Reader votes and developer ratings from the same platform were grouped into one category. A paper with high scores from both readers and developers then counted as only one signal, and it fell below the threshold for selection.

The AI pointed out that if the populations casting votes differ, the signals are separate. Readers and developers are distinct populations even when they use the same platform. This correction was possible because the definition already contained the word "independent." Without it, there would have been no basis for deciding what counts as independent and what does not.

Anthropic's developer documentation recommends defining success criteria before refining any instruction to an AI. A criterion is, in effect, a definition written in a form that allows measurement.

04A definition here means a set of conditions that determine inclusion or exclusion

The kind of definition discussed in this article is not a dictionary entry. It is a combination of conditions that determines whether a given item qualifies or does not qualify.

Having AI write a definition means asking it to lay out these four elements, then reviewing each condition against your own purpose. The definition the AI returns is a starting point, not a final answer.

05Defining review items as conditions keeps every reviewer on the same axis

In pharmaceutical promotional material review, the same principle applies directly. When the items to check during a review are listed in broad terms such as "appropriateness of expression," each reviewer ends up covering a different scope. When the condition is written as "any statement that exceeds the approved indications," inclusion or exclusion can be decided consistently.

The same gap appears when classifying past review findings. A category labelled "expression correction" puts together corrections for overclaiming and corrections for readability. Adding a condition that separates "inconsistency with approved matters" from "readability improvement" makes the classification reproducible, whether done by a person or by a machine.

When AI is asked to classify past findings, the result depends on whether a definition is given in advance. Without one, the AI invents its own axis. That axis may not match the purpose of the work. With a definition provided first, the AI's classification and the reviewer's judgement sit on the same axis.

06AI can draft a definition, but choosing the axis is the person's job

AI can produce definitions because language models have learned, from a large volume of text, where the boundaries of a concept tend to fall. They hold a statistical map of the contexts in which a term is used and the contexts in which it is not. This is why a request such as "define hot topic" returns a list of conditions.

The limitation is that the definition the AI returns leans toward whichever usage appears most often in its training data. In the operator's case, the AI initially split categories by data source, which is common in the research literature. It did not match the operator's purpose. Only when the operator asked whether readers and developers are the same population did the axis shift to one based on the responding group.

Anthropic's prompting guidance states that the more specific an instruction is, the fewer corrections are needed afterwards. A definition is one way to make an instruction specific: replacing an ambiguous term with a set of conditions narrows the range within which the AI has to guess.

The first definition the AI returns is not necessarily the best one. Asking what changes when the axis changes, and then choosing the axis that fits the purpose, is work that stays with the person.

07From tomorrow: three questions to build a definition before collecting

Before asking AI to collect information, pose three questions.

1. "What are the conditions for inclusion and exclusion?" The AI returns a list of conditions that separate qualifying items from non-qualifying ones.
2. "How would the result change if the classification axis were different?" Ask what happens when categories are split by source versus by population, by time period, or by another dimension. Candidate axes appear.
3. "Is there anything important that this definition would miss?" This checks whether the exclusion boundary is set correctly.

Read the AI's answers to all three, settle the conditions and the axis, and only then move on to the collection programme or the analysis instruction. Keep the definition as a written document. The next time a similar type of information needs to be collected, the definition ensures the same criteria apply.