For people who work in a research or policy unit, do research, analyse data or prepare documents for decision-makers: country briefs, studies and reports, data sheets and surveys, sources that do not match, monitoring a topic, documents in another language. Nine cases, each with the request to copy and adapt and with what to check before a figure leaves your work.
The examples use invented documents and data, except the first, run on a real country with public sources. Each case lists the stages of the manual that teach the move. The cases were tested in Claude in the Italian edition of this guide, and the cases that report what the model did in a test were tested again in English on these documents, in November 2026. Where a test showed a typical mistake, the case says so.
Where the test used a fictional document, the case offers it for download: you can redo the test and compare the answer with what the case describes.
Before you start: every figure has a date, a unit and a source
Research lives on figures taken from others, and the model is very good at writing them up neatly even when it does not have them. Four rules apply to every case below.
- Every figure carries the year or month it refers to, the unit and the source with the link. Write it in the request, every time or in the project instructions: a 2024 figure next to a 2026 forecast, with no dates, looks like a comparison and is not one.
- The calculations are done by code, not by the model in its head (stage 7). Means, percentages, correlations: ask for them to be run with a program and to see the calculation.
- When it searches, the model must open the pages it cites. When it remembers, it can invent plausible titles, authors and figures. Active search makes the difference (stage 15), together with a request for a link for every source.
- Microdata, survey answers, interviews: remove names and identifiers before uploading them to a personal tool. The analysis needs the rows, not who wrote them.
1. The country brief
The situation. You follow several countries and keep a brief for each, with the same sections: political picture, economy, external accounts, relations with the UK, risks. It has to be redone every time something moves.
The request. You set it up once: a project (stage 12) with the format of the brief, the list of sources you accept and two old briefs as examples. In the instructions: every figure with the period, the source and the link; if two sources do not match, both with their date; "not found" where a figure is missing. Then, with deep research: "Country brief: [country]. Format and sources from the project." A month later: "Compare today's brief with the one of [date]: only the figures that changed, before and after, with the source."
What to check. The periods: every figure with its month or quarter, because the search returns figures of different years side by side. In the test, on Türkiye, the brief put the official inflation figure next to an independent estimate about fifteen points higher, both dated, wrote "not found" where a figure was missing, and said which figures came from a data aggregator rather than the official release: those are the ones to open on the source. One economic figure opened on the source's table. The risks section, sentence by sentence: it is the part the model writes best and cites least.
The full walkthrough. Run for real on Türkiye on 31 October 2026, with the brief from the test to download: The researcher: the country brief. An hour and a half the first time, half an hour for each country afterwards.
Stages. 12 (the first project), 15 (when it searches instead of remembering).
2. What the studies say on a topic
The situation. You need to know what recent studies say on a precise question, before you write or decide.
The request. With web search switched on:
"Find five studies published since 2023 on [question]. For each: authors, year, journal or publisher, the link to the study's page, the method in one line, the main result in one line with the figure. Only studies whose page you opened in this search: no citations from memory. At the end, where the studies disagree."
What to check. That every study exists: open the links, and for at least two search for the title in a search engine. A reference with real authors and an invented title is the typical mistake of models answering from memory. That the result reported is the study's and not that of a newspaper article about it. The figure: look for it in the text of the study. Look also at what it opened: in the test the model said for each study whether it was a peer-reviewed article, a university press release or a report, and for the two pages that would not open it said so and named the secondary pages it had used instead. Those are the figures to check first. That is the behaviour to ask for.
3. The figures in a long report
The situation. A hundred-page report, and you need ten figures from it, knowing for certain where each one is.
The request. Upload the report, then:
"Extract the main figures in a table: indicator, value, unit, year or period, page. Separate data from forecasts. If the same indicator appears with different values in different places in the report, list them all with the page and flag it."
To redo the test. The fictional document used: case3_observatory_report.txt.
What to check. Internal inconsistencies: the summary and the table of the same report sometimes disagree, and whoever copies from the summary copies the mistake. The units: thousands or units, per cent or percentage points. The years: a figure from the year before, quoted because it is the latest available, easily ends up in the row for the following year.
Stages. 3 (a document of your own), 4.
4. A data sheet, before the analysis
The situation. You have a data file to analyse: means, comparisons, a correlation, a chart.
The request. Upload the file, then:
"Before calculating, check the data: missing values and how they are coded, out-of-range values, duplicate rows. Tell me what you found and what you propose to do about it, and wait for my go-ahead. Then calculate with a program the descriptive statistics of each column and the correlation between [column] and [column], with the chart. Show me the code."
To redo the test. The fictional document used: case4_local_authorities.csv.
What to check. The first step, which is the one that matters: missing values written as -99 or 0, which enter the means and distort them; a figure with one zero too many, which on its own moves a correlation. The decisions on these values are yours, and they must be written into the work. In the test the three bad values in thirty rows took the correlation from 0.91 to 0.09; the model found them, stopped, and proposed excluding the typo instead of correcting it, because the correction would have been its own guess. The code: if you do not read it, ask for every step to be explained in plain English.
Stages. 7 (Excel: formulas, data and charts), 4.
5. When the sources do not match
The situation. Three sources, three different figures on the same thing. Which one do you use?
The request. Paste the three passages with where they come from, then:
"These sources give different figures on the same topic. For each: what it measures exactly (unit, population, period), who produces it, when. Then tell me whether the figures really contradict each other or measure different things, and which source you would use for which statement. Do not average them."
To redo the test. The fictional document used: case5_three_poverty_sources.txt.
What to check. The definition: households or people, residents or citizens, calendar year, financial year or the last twelve months. Two "different" figures are often both right. A source with no date or no unit is not used, even if its figure is the most quoted.
Stages. 2 (the same request, done well), 15.
6. The results of a survey
The situation. Survey answers in a spreadsheet, to be turned into percentages and comparisons between groups.
The request. Upload the file, then:
"Calculate with a program the percentage of satisfied respondents (answers 4 and 5) overall and by region. For each percentage, the base: how many valid answers are behind it. Tell me how you treat blank answers. Flag the groups with fewer than ten answers."
To redo the test. The fictional document used: case6_survey.csv.
What to check. The base: the percentage changes depending on whether blank answers are in the denominator or not, and a serious report says which choice it made. Small groups: 40 per cent of five answers is two people, and it cannot stand a comparison with the other regions.
7. The note for the decision-maker
The situation. You have read the report. Your director has five minutes.
The request. In the same chat as case 3, or with the report uploaded:
"Write a one-page note for the director: what the report says in five points with the figures, what it does not say or does not allow us to say, three questions to ask the authors. Call forecasts forecasts, with their assumptions."
What to check. That the inconsistencies found in case 3 stay in the note: a polished summary tends to pick one of the two figures and leave out the other. Forecasts: without the assumption behind them they become promises. The section "what it does not say": it is the part that makes an analytical note different from a summary.
Stages. 2, 5 (the finished work).
8. Monitoring a topic
The situation. You follow a topic that moves every week: a bill going through Parliament, a market, a sector. Every Monday you run the same searches again.
The request. The first time by hand, with web search switched on:
"Search for what is new in the last seven days on [topic], only from these sources: [list with the addresses]. For each item: date, what happened in two lines, link. Today is [date]. If there is nothing in the week, say so: do not fill the gap with older news."
When the result convinces you, the same request becomes a task that comes back on its own every Monday (stage 20).
What to check. The dates: news from three weeks ago, picked up by a site yesterday, easily slips into the week. "Nothing new" is a right answer, and the model tends to avoid it. A link for every item, opened. And ask what it could not check: in the test it found only one new item in the week and said which sources it had not been able to go through fully. Without that line, "nothing new" cannot be told apart from "I did not look".
Stages. 15, 20 (the task that comes back on its own).
9. A document in another language
The situation. The document you need is in German or French, and the summary has to be written in English for people who do not read it.
The request. Upload the document, then:
"Summarise this document in English in ten lines, with all the figures. Then make a glossary of the technical terms with the English rendering you chose. Watch out for false friends and number separators: flag the places where a literal translation would be wrong."
To redo the test. The fictional document used: case9_german_country_brief.txt.
What to check. The figures: in German and French the comma is the decimal point and the point (or a space) separates thousands, so "3.200" is three thousand two hundred, not 3.2, and "1,5 Milliarden" is 1.5 billion. A German "Billion" is an English trillion, and a French "milliard" is a billion. The false friends: German "eventuell" and French "éventuellement" mean "possibly", not "eventually"; "aktuell" and "actuel" mean "current", not "actual"; "sensibel" and "sensible" mean "sensitive", not "sensible". "Prozentpunkte" are percentage points, not per cent. The glossary is useful afterwards too: if you often work from the same organisation's documents, put it in a project's instructions.
Cases tested on 7 November 2026.