# How to use this file
AI for Data and Documents — TOPIC 3
Reading Documents at Scale
Summarize, extract, compare, and classify policy documents

HOW TO USE THIS FILE
  1. From the "Policy Documents" page, pick ONE team package and download its
     documents. Every team's files are named the same way:
        Doc1, Doc2, Doc3   (some teams have more)
     so these prompts work for any team.
  2. Open a new chat at claude.ai and UPLOAD the documents for your team.
  3. Copy a prompt, paste it into Claude, and run it.
  4. Keep the AI grounded in the source: when it makes a claim, ask for the quote.

THE GROUNDING RULE (use it constantly)
  AI text tools can "hallucinate" — invent plausible-sounding facts. For every
  factual claim, require a short quotation and where it came from:
        "Support each point with a short exact quote from the document and name
         which document (Doc1, Doc2, ...) it came from. If the document does not
         say, reply 'not stated' rather than guessing."

---

# Faithful summary of one document
I have uploaded Doc1. Summarize Doc1 in 5 to 7 bullet points.

Rules:
  - Use ONLY what is in Doc1. Do not add outside knowledge.
  - After each bullet, put a short exact quote (under 20 words) in quotation
    marks that supports it.
  - If something important is unclear or not stated, say "not stated" rather
    than filling the gap.
Then give me a one-sentence takeaway of the document's overall message.

---

# Extract structured facts into a table
From Doc1, extract the following into a clean table with columns
[Item | Value | Exact quote]:
  - the issuing organization (e.g., IMF, World Bank, a central bank)
  - the country or region the document is about
  - the date or time period
  - any headline economic figures mentioned (GDP growth, inflation, policy/interest
    rate, unemployment, debt) — one row per figure
  - the main policy recommendation or decision, if any

If a field is not present in the document, write "not stated". Do not invent
numbers. Show the table.

---

# Classify the stance / topic
Read Doc1 and classify it on three dimensions. For each, give the label AND a
one-line justification with a short quote:

  A. Document type: [central-bank statement / IMF assessment / World Bank report /
     rating action / company or market report / communiqué / other]
  B. Overall economic outlook: [positive / mixed / negative / neutral]
  C. Policy direction it points to: [tighten / ease / hold / not applicable]

If the document does not give enough basis for a dimension, answer "cannot tell
from this document".

---

# Compare several documents at once
I have uploaded Doc1, Doc2, and Doc3 for my team. Build ONE comparison table with
one column per document and these rows:
  - Issuing organization
  - Country / region and time period
  - Headline growth figure (or "not stated")
  - Headline inflation figure (or "not stated")
  - Overall outlook (positive / mixed / negative)
  - Main message in <= 12 words

Fill every cell using only the documents. After the table, write two sentences on
the single biggest point of AGREEMENT and the biggest point of DISAGREEMENT across
the documents, each backed by a short quote naming the document.

---

# Find the numbers and put them side by side
Across all the documents I uploaded (Doc1, Doc2, ...), pull out every explicit
economic figure (growth, inflation, interest/policy rate, unemployment,
debt/deficit). Present them in a table with columns
[Document | Indicator | Value | Period | Exact quote].

Do NOT compute or estimate any figure that is not written in a document. If two
documents give different numbers for the same indicator, keep both rows so I can
see the disagreement.

---

# A short grounded brief
Using only my team's documents, write a half-page brief (plain English) that a
busy reader could skim, covering: who issued these documents and about where/when;
what they say about growth and inflation; and where they agree or diverge. Every
factual claim must carry a short quote and the document name. End with one line:
"What these documents do NOT tell us: ___" to make the limits explicit.

---

# VERIFICATION HABITS (Topic 3)
- Demand a quote for every fact. No quote -> treat it as unverified.
  - Watch for invented precision (e.g., a rate with decimals the source never
    gave). Ask: "Which document and sentence is this from?"
  - If the model summarizes something your document does not contain, that is a
    hallucination — call it out and re-prompt with "use only the uploaded text."
  - Spot-check two or three quotes by finding them yourself in the PDF/text.