23 Using an assistant for a literature search

What people usually ask

Find recent papers on machine learning for membrane fouling prediction.

What comes back

A list of review titles with plausible authors, two papers that do not exist, and nothing on any particular membrane or feed.

Ask this instead

First:

Help me plan a literature search on machine learning for predicting fouling in <membrane type> membranes treating <feed>. Do not give me papers.

Start with the mechanism, in plain words and under 150 words: what fouls these membranes with this feed, what is measured while it happens, and what a model would be asked to predict.
Then query sets grouped by theme (the fouling mechanism, the measurement, the model type), four to six queries each. Every query names a specific membrane material or module and a specific feed; no query may say "membrane" or "wastewater" on its own. Use the boolean syntax of <database>. Peer reviewed journals and conferences only.
Save the mechanism and the query sets as plan.md. I will run the queries, download the papers and judge them myself.

Then, with a downloaded paper:

Attached is one paper from the search. Against plan.md, say which query found it, which theme it serves, what it reports that the plan did not expect, and whether the plan should change. One paper per message; do not summarize the paper.

Attach: Nothing at first; the downloaded papers later, one per message.

Why it works: the assistant builds the queries and the database returns the papers, so nothing enters the plan that was not downloaded and read.