Testing an in-depth research officer on a real literature review

What was done

Aaron Schiff recounts (March 2025) having spent 30 minutes writing his prompt. By checking, he found unverifiable results, unreliable sources (including a blog post and a Facebook publication), missed references and in-depth research, without access to paid publications. Some of the results remained usable. His conclusion: "at the limit of useful", a starting point and not a customer deliverable.

Before you start

Tool cited by the source: ChatGPT Deep Research. Fill in the bracketed items: [sujet], [pays / période]. Strip confidential data that is not needed and check your tool’s access terms. Features and plans can change.

The method, in three steps.

1. 01

   Gather the context

   Spot the bracketed items in the prompt ([sujet], [pays / période]) and prepare what they ask for.

2. 02

   Adapt the prompt

   Copy the prompt below, replace the bracketed items, then paste it into ChatGPT Deep Research.

3. 03

   Reread before using

   Reread the result, correct it if needed and keep the final decision. Caveats are detailed below.

Your starting point.

Replace the bracketed items, then adapt the proposal to your situation. This prompt is written by the iadujour desk It is not a quotation from the source.

Review published estimates on [sujet] for [pays / période].

For each estimate: value, unit, year, method, population studied, complete reference and link.
Rules:
- only academic, institutional or press sources; no blogs or social networks
- goes back to primary sources when a report quotes other sources
- indicates the studies you could not read (paying)
Finished with the list of tracks to explore by hand.

What deserves your attention