From Raw Documents to Reusable Insight

Open Notebook becomes more valuable when research does not end with a single answer. This chapter focuses on the step between retrieval and lasting understanding: using transformations to extract comparable insights, then turning selected results into notes that can be edited, searched, and reused. The practical goal is not to automate judgment away, but to make repeated analysis consistent enough that a researcher can spend more attention on differences, exceptions, and unanswered questions.

A transformation is a repeatable question

Chat is useful when a researcher wants a conversation. Ask is useful when they want the system to search and synthesize around one broad question. A transformation begins with a different need: the researcher wants to ask essentially the same analytical question of many sources and receive outputs that can be compared.

Imagine ten research papers about a common problem. Asking “What is this paper about?” ten times in Chat might produce ten answers with different emphases, lengths, and structures. One paper might mention its method first, another its conclusion, and a third might omit limitations entirely. The responses could still be individually fluent, but they would be difficult to compare. A transformation lets the researcher define a reusable extraction such as the research question, method, key findings, limitations, and relevance, then apply that pattern across the collection.

The value comes from repeatability, not from the template being inherently correct. A poor template applied consistently produces a consistently incomplete view. If the researcher asks only for findings, the resulting insights may omit sample limitations or the conditions under which the findings hold. If the template asks for too many fields, the model may compress each one until the output is technically complete but practically thin.

A useful transformation therefore starts with the decision the researcher expects to make. For a literature review, the relevant fields may be method, population, finding, limitation, and connection to the review question. For interview material, the fields might be decision, evidence, unresolved concern, and next action. For technical documentation, the useful structure could be purpose, assumptions, failure conditions, and operational consequence. The template should reflect the source type and the comparison that matters, rather than imitate a generic summary form.

The template is strongest when each field has a boundary. “Finding” might mean what the authors report, while “interpretation” might mean what the researcher infers. “Limitation” might include only caveats acknowledged by the source, or it might invite a separate critical reading; those are different tasks and should not share an ambiguous label. “Evidence” might request a page, section, quotation, or timestamp, depending on the input. When a field is not present, the template can ask the model to say so. That is more useful than rewarding it for filling every blank.

A literature-review transformation, for example, could ask for five short sections: the question the paper addresses, the method it uses, the result it reports, the limitation it acknowledges, and the relevance to a particular review. It could request a page or section reference for the result and limitation, and ask the model to separate the paper's claim from the researcher's possible interpretation. Applied to several papers, the output would not settle which study is strongest. It would give the researcher a comparable starting surface on which differences in sample, method, and qualification can be inspected.

This is the difference between a template that reduces work and one that merely reduces thought. A good template makes the next judgment easier by exposing the variables that matter. A bad one produces uniform prose that hides whether the sources answered the same question, used comparable evidence, or left important fields unaddressed. Repeatability is valuable only when the repeated operation preserves the distinctions the decision depends on.

The prompt should make its requested structure visible. “Summarize this” leaves the model to decide what summary means.