Where an assistant belongs in research
An assistant can sit in a reading workflow in a few narrow places. It can help phrase a question, list terms you already suspect, or compress a document you have open. It cannot replace the document. This page draws that line for study. It does not recommend a product, and it does not suggest that a faster draft leads to a better financial decision.
The confusion usually starts with the word “research.” In a newsroom or a classroom, research means going to sources. In a product label, the same word can mean “the model wrote a long answer.” Those are different jobs. Only one of them leaves a trail you can check.
A draft is not a finding
Using a model to draft a question is ordinary. “What does this statistical office mean by residence?” is a better starting point than a vague worry. The draft can be clumsy and still useful, because you will edit it against a real glossary. The finding, if there is one, begins when the glossary answers. The draft did not answer.
The same split applies to an outline. An outline proposes an order of topics. Order is a writing choice. It is not evidence that the first topic matters more in the world, and it is not a ranking of companies, sectors, or instruments. If an outline starts to sound like a plan of action, it has left the job of drafting.
- Draft
- Text you expect to change after you read a source.
- Source
- A document with an author, a date, and a method you can inspect.
- Trail
- The path from a sentence back to that document. A continuation alone has no trail.
Confirmation stays with the source
Suppose the assistant names a report. The next step is to open the report, not to ask the assistant if the report is real. Models can name plausible titles. A plausible title is not a publication. When the file is open, read the definition in the file. If the assistant’s sentence and the file disagree, the file wins.
Citations added after the fact need the same check. A link can point at the right site and the wrong paragraph. A quotation can be stitched from two passages. The workflow is short: open, find the passage, compare. Skipping the compare step is how a fluent paragraph acquires false footnotes.
Summaries need the text in hand
A summary of a document you uploaded, or of a passage you pasted, is a compression of that text. It can drop a condition, a date, or a denominator. Read the summary beside the passage, not instead of it. A summary of “the topic” with no passage attached is a new composition. Treat it as uncited prose.
Numbers are the easiest thing to warp in compression. A rate and a level are different. A revised figure and a first release are different. If the summary keeps a number, the source of the number should still be one click away. This is a reading rule, not a claim about any particular statistic’s direction.
Where the assistant stops
It stops before a personal decision. A question about what a term means can sit in the workflow. A question about what you should buy, sell, or hold cannot be answered by a cleaner paragraph. The paragraph does not know your situation, and this site will not pretend that an assistant does.
It also stops before authorship. If you publish a sentence, you are responsible for the trail behind it. “The model wrote it” is not a source line. The source line names the document. If there is no document, say that the sentence is unverified. That honesty is part of the workflow, not a flaw in it.
- Use the assistant to phrase, not to certify.
- Open every named report before you repeat it.
- Keep decisions about money outside the draft.
Close
An assistant belongs where the work is still a draft and the source is about to be opened. It does not belong where a sentence needs a trail and none is being checked. The line is procedural. It is not a product review.