Page 90 - CyberSecurityEssentialsEnglish

The context window is the amount of text, images, and conversation the model can refer to
within a single request, and it usually includes both the input and the output. A larger
window allows longer files and conversations, but it does not guarantee that every piece of

information will be retrieved accurately. Research on the “lost in the middle” effect shows
that performance can decline when the important information sits in the middle of a long

context, even in models that support large inputs [13].

Begin with a map of the document: the title, the version, the sections, the pages, the

appendices, and what you are looking for. If the task requires comparing several documents,
give each one an identifier, a source, and a date. Ask the system to state the document
identifier and the page or the quotation with every inference. This does not guarantee

correctness, but it makes tracing possible.

Chunk the material by meaning: by chapter, topic, period, or record type. Mechanical

chunking every ten pages may separate a table from its explanation. Use a uniform
extraction template for each chunk, then assemble the results into a matrix. Do not ask for
a different summary of each part, because differing fields make aggregation harder.

Maintain an updated working brief that contains the objective, the decisions, the definitions,
the assumptions, the sources, and what remains open. This is an organizing term the book
uses to reduce the loss of state between rounds. It must stay short and reviewable, rather

than turning into another long document that repeats everything.

In long conversations, do not keep adding corrections on top of old instructions. When the
direction changes, write a clean brief and begin a new stage. Keep important versions

outside the conversation. Tools also differ in how they manage or truncate context, so test
the actual task rather than relying on the advertised number alone [4][13].

Progressive Outlining and Layered Aggregation
On a long task, begin with an outline rather than a full draft. Ask for the section headings,
the source required for each section, and the open questions. Review the outline, then write

one section or one unit. This approach is called progressive outlining here; it is an organizing
practice rather than a technical device.

When summarizing dozens of documents, use layered aggregation: uniform extraction from

each document, then grouping by topic, then a final synthesis. Keep the quotations and page
numbers beside each conclusion. Do not ask the system to read everything and write the
final conclusion directly, because you will lose the evidence trail and it becomes hard to find
where the error occurred.

Use verification samples at every stage. Review part of the extraction by hand, then part of
the aggregation, then the final conclusion. If an error appears early, fix the template before

89