# How to use this file
AI for Data and Documents — TOPIC 4
From Documents to Analysis
Sentiment & tone, change over time, and evidence-based synthesis

HOW TO USE THIS FILE
  1. Use the SAME team package you used in Topic 3 (files named Doc1, Doc2, ...).
  2. Open a new chat at claude.ai and UPLOAD your team's documents.
  3. Run the prompts in order. Topic 4 goes deeper than Topic 3: it measures tone,
     tracks change, and builds a synthesis — always with citations.

THE CITATION RULE (keep it on the whole session)
  "Every claim must cite the document (Doc1, Doc2, ...) and include a short exact
   quote. If the documents do not support a claim, say so instead of guessing."

---

# Tone / sentiment of one document
Analyze the TONE of Doc1 toward the economy.

  1. Give an overall tone label on this scale:
        very negative / negative / neutral / positive / very positive
  2. Separately, if this is a monetary-policy or central-bank style document,
     rate its policy tone as: hawkish / neutral / dovish
     (hawkish = leaning toward higher rates / fighting inflation;
      dovish = leaning toward lower rates / supporting growth).
  3. Justify BOTH labels with three short exact quotes that carry the tone (the
     specific words that signal it, e.g. "elevated", "attentive to risks",
     "solid pace", "uncertainty").
If a dimension does not apply to this document, say "not applicable".

---

# A simple, transparent tone score
Build a transparent tone score for Doc1 that I can check by hand.

  1. Count how many times clearly HAWKISH / negative-for-easing words appear
     (e.g., inflation, elevated, overheating, tighten, attentive to risks,
     price pressures).
  2. Count clearly DOVISH / supportive-of-easing words (e.g., slowdown, weak,
     downside risks, support, accommodation, cut, ease).
  3. Report both counts, list which words you counted, and give a net score
     (hawkish minus dovish). 
State plainly that this is a simple word-count proxy, not a sophisticated model,
and show the word lists so I can audit them.

---

# Track tone/stance over time or across sources
I have uploaded several documents (Doc1, Doc2, Doc3, ...) that differ by date or
by source. Put them in the right order (use the dates inside the documents; if a
date is missing, say so).

Then build a table: [Document | Date or source | Tone label | Policy stance |
one key quote]. After the table, describe the TREND in two or three sentences:
is the tone getting more hawkish or more dovish, more positive or more negative,
as we move through the documents? Base every statement on the quotes.

---

# How did the LANGUAGE change?
Compare the earliest and the latest document in my set. Identify specific PHRASES
that were added, dropped, or changed between them (central banks and the IMF
signal shifts by changing a few words). Present:
  [Phrase in earlier doc] -> [Phrase in later doc] -> [what the change signals]
Use exact quotes from each document. If a phrase is unchanged but important, note
that too. Do not speculate beyond what the wording supports.

---

# Evidence-based synthesis with citations
Write a one-page synthesis that answers: "What is the overall economic picture
and policy direction described by my team's documents?"

Requirements:
  - Organize it as: (1) Growth, (2) Inflation / prices, (3) Policy direction,
    (4) Risks.
  - EVERY sentence that states a fact must end with a citation in the form
    (Doc2: "short quote").
  - Where documents disagree, present both sides with their quotes.
  - Finish with a "Confidence & gaps" paragraph: what is well supported, what is
    thin, and what the documents do not address.

---

# Turn the synthesis into a defensible claim check
From your synthesis, list the 5 most important CLAIMS as a checklist. For each,
show: the claim, the document(s) and exact quote(s) that support it, and a
confidence level (high / medium / low) with one line of reasoning. This is the
artifact a careful analyst keeps so every statement can be traced to a source.

---

# VERIFICATION HABITS (Topic 4)
- A tone label with no quotes is an opinion, not evidence — always demand the
    words that justify it.
  - For the word-count score, read the word lists: are they fair? Add or remove
    words and re-run to see how sensitive the score is.
  - Check that "change over time" claims use the actual dates inside the
    documents, not the model's assumptions.
  - Trace at least two citations back to the source yourself. If a quote is not
    in the document, discard the claim and re-prompt.