The researcher: the country brief

Session: an hour and a half the first time, half an hour for each country afterwards. Time to see the result: a month, when you redo the brief and compare. You need: the format of the brief your office uses, even just as a list of sections; the list of sources you accept; deep research (stage 15) and a project (stage 12).

People who follow several countries do a job that is repetitive and delicate at once: the same brief, with the same sections, on different realities, to update every time something moves. The model produces it in ten minutes; the problem is that it produces it even when the data isn't there, mixing years and sources with the same confidence. Today you build the set-up that prevents it: a fixed format, a boundary of sources, and the comparison over time. The example is Türkiye, with a search made on 31 October 2026.

1. The format is yours

Write in a file the sections of the brief, in the order you want them: institutional and political picture (who governs, since when, next deadlines); economy (growth, inflation, public finances, debt, with the latest figure and the forecast); external sector (balance of payments, reserves, currency, ratings from the three agencies); relations with your country and with the European Union (trade, presence, agreements); risks and things to watch in the next six months. For every section the rule of form: every figure with the year or month it refers to, the source and the link; never an average of different sources; "not found" where it is missing, and "not found" also when the figure is there but the page does not say which year it refers to.

2. The sources, chosen beforehand

It is step 3 of stage 15, on this work: the list of sites you accept data from, one per line with the address. For the figures, the institutions that produce them (the IMF, the World Bank, the OECD, the country's statistical office, the central bank); for the political picture and the risks, the sources your office already uses; for your own country, the trade data and the ministries' briefs. You decide which sources count; the model uses them. Outside the list it does not search.

3. The "Country briefs" project

Open a project (stage 12) containing the format, the list of sources, two old briefs done by hand as examples of tone and length, and the instructions: "Produce country briefs in the format of the file "Format". Use only the sources in the file "Sources". Every figure with the date it refers to, the source and the link; if two sources do not match, report both with the date of each; if a figure is not there, write "not found". No adjectives, no forecasts of your own: the forecasts are the sources', with the date they were made." If in your tool deep research can be launched from a chat opened inside the project, stay there; otherwise bring format and sources into the request.

4. The first brief

With deep research switched on (stage 15, step 4): "Country brief: Türkiye. Format and sources from the project. For every figure the latest available, with the period it refers to. At the end, the list of sources consulted with the link and the date of consultation." Read it from the bottom, as in stage 15: first the sources, then the figures that would change a decision, then a sample of the rest. In the test, the search on September inflation returned 61 per cent, 49, 33 and 29.7 all together: they are the Septembers of 2023, 2024, 2025 and 2026, and the right figure for the brief was the last one. Without the rule "every figure with its year" one of the four would have gone in at random. The corrections you make become lines in the project's instructions, not corrections to the brief. The brief from the test, with every figure dated, the sources and the "not found" entries in their place, can be downloaded from librovivoia.it/prova/turkey.

5. The second brief, and the comparison

Redo the brief a month later, or do a second country right away. Then: "Compare today's brief with the one of [date]: list only the figures that changed, with the value before, the value after and the source reporting the new one. Do not comment." That list is the real work of anyone following a country: the things that moved. In the test the Turkish central bank's policy rate was 37 per cent until 22 October 2026 and 36 from the day after: a brief made on the 21st and read on the 23rd was already old on that figure, and that is why the brief carries its date at the top. Save it in the project with the date in the name.

6. The check

Two per brief: one economic figure opened on the original source's table, to see that the period is the one declared; and the risks section read asking where each sentence comes from, because it is the part where the model writes best and cites least. When two good sources do not match, it is not necessarily a mistake: in the test the IMF estimated in February 2026 a growth of 4.1 per cent for 2025, and the Turkish statistical office in August published it at 3.7; the brief reports both, with the dates. On the political picture, the date of the next deadlines is always verified: it is the figure that ages first.

What you just saw

Format, boundary, project, comparison over time: stages 12, 14 and 15 on one job. Deep research pays as much as the boundary is tight, and the brief holds over time because it is the project that remembers the format and the sources, not you at every request. The specific risk of this work is the mix of periods: a 2024 figure next to a 2026 forecast on the same line, without anyone saying so; the rule "every figure with its date" is there for that, and in the test it did its work four times in an hour. Chapter 5 describes external memory and search; chapter 9 says which activities bear delegation and who checks.

In the book. Chapter 5, section 5.3 "The memory of LLMs", page 98; chapter 9, sections 9.3 "Which activities stand up to delegation", page 142, and 9.5 "Who verifies, and what", page 146. The book on Amazon.