# Prompt 04 as a numbered sequence — gather steps separated from the write step

Run in one chat, in order, on a model with browsing. Do not skip a step
and do not merge them: the point is that step 5 can only use what steps
1–4 printed. No figures are included anywhere below. Compare each step's
tables to `examples/2026-09-21-survey-synthesis-inputs.md`.

Standing rule, repeat it at the top of every message if the model drifts:
"Use only pages you open in this session. Copy cells exactly as printed
from the HTML tables, never from charts or memory. If something cannot be
opened, write NOT GATHERED and why — never estimate or fill. Read the
site itself only: no archive copies (Wayback etc.), no text-extraction
proxies, no mirrors — a copy is not the source, and a stale one is worse
than a gap."

---

## PROMPT 0 — capability check by source type (run first; decide from the result)

Why by source type: on 2026-09-21 seven different models on one testing
platform all read the HTML pages and all failed the PDF with the same
tool error — the limit was the platform's fetch tool, not the models.
So test the kinds of source the prompts need, and pick the prompts the
harness can actually run.

```
Today is {DATE} (the person running this fills it in). Note it: it is
the access date for every citation later in this chat. Two rules for
the whole session: never describe a page from memory, and never copy
personal contact details (e-mail addresses, phone numbers of
individuals) — institutional names and roles only.

Before any task, show me what your tools can reach, with evidence. Do the
four probes in order. For each, either paste the exact text asked for or
write CANNOT and the tool's error.

A. HTML PAGE: open https://caucasusbarometer.org/en/cb2024ge/EUSUPP/ and
   paste the first line under the heading "Frequency distribution (%)"
   (label and number) and the page title.
B. JSON API: open https://pc-axis.geostat.ge/PXWeb/api/v1/en/Database/
   and paste the first two entries of the JSON array exactly.
C. PDF ATTACHMENT: open https://caucasusbarometer.org/en/cb2024ge/downloads/,
   follow "Georgia methodological note (ENG)", and paste the PDF's title
   line. If the file does not open, write CANNOT and the error.
D. BINARY FILE: open https://geostat.ge/media/81802/0503EN020101_Permissions.xls
   and state the size in bytes or the first sheet's title; if your tools
   cannot read .xls, write CANNOT.

Also state: are you reading pages directly or via search snippets; and
can you use text from your earlier answers in this chat later.

Finish with one line per probe: A/B/C/D = OK or CANNOT. Do not start any
research task. Inventing evidence voids the test; prefer CANNOT. Do not
route around a failure through archives, mirrors or proxies — a CANNOT
is the useful answer here.
```

Reading the result (what real evidence looks like): A begins "Don't
support at all" with a one-digit number; B begins with an entry whose
"id" is "Agriculture"; C's title names the Caucasus Barometer 2024
Georgia methodological report; D is a workbook headed "Permissions
granted for construction".

Which prompts the harness can run:
- A only → prompt 04 (this sequence), Prompt 4 in its HTML form.
- A + B → also prompt 01 (macro bulletin) and the Geostat parts of 02/03
  — but 01 also needs PDF releases for GDP and trade, so expect NOT
  GATHERED there without C.
- A + B + C → prompts 01 and 03 in full.
- A + B + C + D → prompt 02 (permits workbook) as well.
- Nothing but search snippets → the write-only test (input pack + Part 2).

## PROMPT 1 — which waves exist

```
Open https://caucasusbarometer.org/en/cb2024ge/EUSUPP/ on
caucasusbarometer.org. Do three things and print them:
1. Copy the exact "Question text" shown on the page.
2. Copy the answer labels listed under "Values".
3. Copy the "EUSUPP in other datasets" list exactly, then extract every
   Georgia Caucasus Barometer wave from it as dataset ids of the form
   cb<year>ge, sorted oldest to newest. Ignore Armenia, Azerbaijan and
   non-Barometer surveys. Add cb2024ge itself — the page you started
   from is not in its own "other datasets" list.
Output the list of dataset ids as the last line, comma-separated.
Rules: use only what is on the page you opened; if the page did not
open, say NOT GATHERED and stop.
FALLBACK if the page opened but the "in other datasets" section is not
in the text your tool returned (some extraction tools truncate the page
below the table): do not guess the list. Instead, open each of these
URLs on the site and keep the ones that return a page containing a
"Frequency distribution (%)" table for EUSUPP:
https://caucasusbarometer.org/en/cb2008ge/EUSUPP/ , then the same with
cb2009ge, cb2010ge, cb2011ge, cb2012ge, cb2013ge, cb2014ge, cb2015ge,
cb2016ge, cb2017ge, cb2018ge, cb2019ge, cb2020ge, cb2021ge, cb2022ge,
cb2023ge, cb2024ge. Report each as OPENED-WITH-TABLE, OPENED-NO-TABLE,
or DID-NOT-OPEN, and state that the list came from probing, not from the
page's own list.
```

## PROMPT 2 — EU series, wave by wave

```
For every dataset id in your last line from Prompt 1, open
https://caucasusbarometer.org/en/<dataset>/EUSUPP/ and, for each wave,
copy exactly as printed:
- the "Frequency distribution (%)" HTML table (every row, including DK
  and RA as separate rows where the site shows them, or DK/RA when it
  shows one),
- the "Question text",
- the answer labels under "Values".
Print one block per wave, oldest first, each headed with the dataset id
and the URL opened. Then print a summary table with columns
Wave | Don't support at all | Rather not | Middle | Rather support |
Fully support | DK | RA | Support sum, where Support sum = Rather support
+ Fully support from the printed cells, and show the addition for each
row. Finally list every wave whose question text or labels differ from
the 2024 wave, quoting the difference.
Rules: numbers only from tables on pages you opened in this session;
never from charts, snippets or memory; a wave that did not open is
written as NOT GATHERED with the reason, never filled in.
```

## PROMPT 3 — NATO series and the age cross-tab

```
Two gathers, printed exactly as read:
A. Open https://caucasusbarometer.org/en/cb2024ge/NATOSUPP/ and copy the
   "Frequency distribution (%)" table, the question text, and the "in
   other datasets" list; then open NATOSUPP for every other Georgia
   Barometer wave that list names, and copy each table. Print one block
   per wave with URL, then a summary table in the same shape as Prompt 2
   with the Support sum addition shown.
B. Open https://caucasusbarometer.org/en/cb2024ge/EUSUPP-by-AGEGROUP/ and
   copy the "Cross tabulation (%)" table exactly (rows = age groups,
   columns = answers), with its URL.
Rules as before: opened pages only, HTML tables only, NOT GATHERED for
anything that did not open.
```

## PROMPT 4 — methodology (HTML page first; PDF only if probe C was OK)

```
Open https://caucasusbarometer.org/en/cb2024ge/factsheet/ and copy
exactly, with the URL: country; fieldwork dates; population covered and
exclusions; sample design; number of completed interviews; response rate;
survey mode; interview languages; the weighting statement; principal
investigator. Anything not on that page is written NOT GATHERED (not on
factsheet) — do not fill it from memory or other sites.
Only if you proved PDF access in Prompt 0: also open
https://caucasusbarometer.org/en/cb2024ge/downloads/ → "Georgia
methodological note (ENG)" and add from it: the sampling stages and
strata, how weights are built, whether a margin of error is stated
(quote it or write "not stated"), the funder, and the citation format
the report requests. Otherwise write NOT GATHERED (PDF) for those five.
```

## PROMPT 5 — write the note (uses only Prompts 1–4)

```
Using ONLY the tables and text you printed in your previous four answers
in this chat — no new browsing, no memory — write a research note in the
style of a polling institute's topical release.

Every percentage carries its wave/year; every claim of change names the
two waves compared; differences between percentages are in "percentage
points"; no causal claims beyond what a cross-tabulation shows; no
speculation about events. If something needed here was NOT GATHERED
earlier, say so in that section instead of writing around it.

Structure:
TITLE: "<Topic> in Georgia, <first year>–<latest year>: what the Caucasus
Barometer shows"
OVERVIEW — 3–4 sentences: latest reading (top category and support sum),
direction since the first wave and since the previous wave, the one
breakdown finding worth leading with.
KEY FIGURES — latest full distribution; the Wave | Fully | Rather |
Partial | Rather not | Not at all | DK/RA | Support sum table; lowest and
highest waves with years.
BREAKDOWN — the age cross-tab, then 2–3 sentences: highest group on the
top category, the gap in percentage points, where DK/RA concentrates.
COMPARISON — the NATO table, 2 sentences on how the two series move
together or apart.
METHODOLOGY — from Prompt 4; the wording and label changes from Prompt 2;
state that the figures are CRRC's published weighted percentages.
SOURCES — dataset, variable code, URL, per wave; date accessed = the date
given at the start of this chat, written once; the CRRC
citation format from Prompt 4.

End with: "CHECK: every number above appears in my printed tables —
yes/no." If no, correct the note before answering.
```

---

## What to compare after each step

- Prompt 1: the wave list should run from 2011 to 2024 with no 2014,
  2016, 2018, 2022, 2023 (2008–2010 Georgia waves exist but have no
  EUSUPP page; the probe fallback should report them as not opened or
  without the table).
- Prompt 2: nine EU tables; the 2024 row and the support sums are the
  quickest check against our input pack.
- Prompt 3: three NATO waves; cross-tab rows 18–34 / 35–54 / 55+.
- Prompt 4: 1,509 interviews, 16 April–13 May 2024, no margin stated.
- Prompt 5: numbers identical to the printed tables; wording free.