More background is not always better background. A request can become harder to manage when it contains old versions, repeated instructions and examples that pull in different directions. Preparing context is an editorial task: decide what the assistant needs for this job, what it should ignore and how much space must remain for the answer.
Begin with the question Suppose you want a short comparison of two fictional onboarding procedures. You have the current procedure, an archived version, six meeting notes and a long conversation about possible improvements. The immediate question is narrow: which steps changed between the approved versions?
The two procedures belong in the first request. The meeting notes may not. If you add every discussion, a proposed improvement can be mistaken for an approved change. Keep a separate list of material you might need later rather than including it by default.
Label versions and authority Give each source a title, date and status. For example, “Procedure A, approved 2 March” and “Procedure B, approved 14 June.” State that the comparison must use these documents and that an absent detail should be reported as absent. This is easier to review than asking the assistant to infer which file is authoritative.
If two sources disagree, describe how you want the conflict handled. You might ask for a separate conflict row rather than a single blended answer. Do not use the newest date as a universal rule if the newer document is merely a draft.
Reserve room for the result The material you submit and the answer both consume capacity in a model workflow. The exact accounting depends on the service and model. A local token count can help you compare prepared text, but it is not a complete prediction of a hosted conversation or its bill.
In a planning exercise, enter your chosen capacity and reserve space for the intended answer. Count the instructions, source text and examples together. If the source barely fits, shorten irrelevant passages or split the task; do not assume that reducing the answer to a few words will produce a useful analysis.
Split at a meaningful boundary For the onboarding example, you could compare preparation steps first and delivery steps second. Keep the same source labels and comparison columns for both passes. Then review the two results for duplicated or missing steps before combining them.
A poor split cuts through a rule and its exception. A better split keeps a section with its heading, definitions and qualifications. When you remove a passage, preserve enough location information to return to the original document.
Check what actually influenced the answer Ask for a source location beside each reported change. Open those locations yourself. If an output cites the archived procedure as current, fix the input labels or the instruction before spending time polishing the prose. A capacity problem and an authority problem require different repairs.
Keep a compact context inventory with the accepted result: included sources, excluded sources, versions and unresolved conflicts. On the next run, update that inventory before reusing the prompt. The aim is not to give the tool everything you know. It is to provide the material needed for a result that another person can trace and inspect.
Keep the inputs and review notes together in the token budget workspace, then save a backup you can reopen.