Predict Intelligence

Predict Intelligence Skill

You generate a professional PDF intelligence brief. The user should grasp the key finding in 30 seconds. You own information efficiency.

Agent Requirements

First-Time Setup

pip install playwright
playwright install chromium

No other Python packages needed (Jinja2, requests, etc. are NOT required).

How It Works

Read template → Do research → Write HTML → Convert to PDF

1. Read SKILL_DIR/templates/report_template.html — your structural reference.
2. Research and analyze the topic (Steps 1–8 below).
3. Generate a NEW .html file following the template's exact structure.
4. Convert to PDF:

   python SKILL_DIR/scripts/to_pdf.py report.html predict_report.pdf
   

The template IS the spec. It contains:

• All CSS (copy verbatim — never modify)
• All D3 visualization code (copy verbatim — only change data variables)
• Example content showing exact formatting for every section
• Extensive comments explaining what each section does and how to fill it

Step 0 — Domain Detection

Classification bar is always: ANYGEN PREDICT INTELLIGENCE ASSESSMENT with #YY-MM-DD on the right (2-digit year, e.g. #26-03-12).

The report structure and design are IDENTICAL across all domains. Only research sources and visualization choices change. See Domain Adaptation at the bottom.

Step 1 — Parse Query

Extract:

• event: what is being predicted
• actors: who is involved
• regions / sectors: geographic or industry scope
• timeframe: any dates or deadlines
• question type: temporal ("when"), binary ("will"), or multi_outcome ("what")

Step 2 — Research

Search the web at least 8 times, covering:

1. Breaking news (last 48 hours)
2. Diplomatic / military / industry developments
3. Decision-makers' public statements
4. Historical precedent
5. Expert / think-tank / analyst reports
6. Regional or sector-specific media
7. Economic / financial context
8. Low-probability high-impact wildcards

CRITICAL: For every fact, save the exact article URL.

For geopolitical / diplomacy / military / sanctions topics, ALSO collect:

• 3–5 key locations (cities, capitals, bases, meeting venues) with lat/lon
• ISO alpha-2 country codes of all parties involved
• Any diplomatic channels, proxy lines, or trade routes between locations This data is required for the V2 Regional Map in Step 7. Collect it now.

fact: "Oman hosted senior US-Iran talks on March 5"
url:  "https://reuters.com/world/middle-east/oman-hosts-rare-us-iran-talks-2026-03-06/"

Homepage URLs are forbidden. Every URL must point to a specific article.

Step 3 — Formulate Predictions

Create outcomes sorted by probability descending. All must sum to ~100%.

The verdict has five parts:

• verdict_number: top probability as text (e.g. "34%") — displayed at 52px, the BIGGEST element
• verdict_outcome: 3–6 word label
• verdict_detail: 1 sentence — "Most likely path: ..."
• verdict_bg (Context): 2–3 sentences of BACKGROUND BRIEFING — what led to the current situation, key prior events, the "前情提要" (previously on...)
• verdict_context: 2–3 sentences of assessment reasoning with key evidence

Calibration:

Step 4 — Key Drivers (5 items, causal logic required)

Each driver MUST have:

• Title: 3–5 words
• Direction: positive (↑ increases likelihood) or negative (↓ decreases)
• Causal logic: 1 sentence: fact → mechanism → directional impact
• Source URL: exact article URL

❌ Bad: "Sanctions pressure — Iran inflation at 45%" ✅ Good: "Sanctions pressure ↑ — 45% inflation pushes Tehran toward concessions for relief"

The bad one states facts. The good one explains the causal chain.

Step 5 — Watch List (5 items, forward-looking only)

Each item is a FUTURE event or trigger that could shift the predict. No past events. No source URLs (these haven't happened yet).

Each item MUST have:

• Date or window: specific date or range (e.g. "Apr 3", "Late March")
• Trigger: 3–6 words describing what might happen
• Conditional impact: 1 sentence — "If [X] → [how probability shifts]"

✅ Good: "Apr 3 — UNSC sanctions vote — If passed → ceasefire probability ↑10-15%" ✅ Good: "Late Mar — FOMC meeting — Dovish tone → ↑ rate cut odds; hawkish → ↓" ❌ Bad: "Mar 5 — Secret Muscat meeting happened" (PAST event — belongs in Key Drivers)

The distinction:

• Key Drivers = WHY the current probability is what it is (past + present evidence, with source URLs)
• Watch List = WHAT could change the probability next (future triggers, no sources needed)

Step 6 — Polymarket Data

Run:

python SKILL_DIR/scripts/fetch_polymarket.py --query "<keywords>" --limit 10

From results:

1. Add your probability estimate for each option
2. Calculate delta = your_estimate − market_probability
3. Select 3 markets with highest absolute delta
4. Sort by delta descending (most undervalued first)

If no script available, search "polymarket [topic]" on the web. If no markets exist, skip the Polymarket section entirely.

Step 7 — Visualization (REQUIRED)

⚠ MANDATORY MAP RULE — READ THIS FIRST

If the topic involves countries, regions, borders, military, diplomacy, or sanctions → you MUST include V2 (Regional Map). Non-negotiable.

Do NOT skip V2 because:

• "I only want one chart" → V2 IS that one chart. Add a second if needed.
• "It needs extra dependencies" → TopoJSON is already in the template <head>. Just keep the <script> tag. Do NOT remove it.
• "I don't have good marker data" → You researched 8+ sources in Step 2. You have city names, country names, and actor locations. Use them. Minimum: 3 markers from your research + highlight the relevant countries.

Only these domains are exempt from V2: financial, technology, corporate, budget/aid (unless they have a geographic dimension).

Choosing visualizations

Pick 1 or 2 types total. Minimum 1, maximum 2.

V2 data — how to fill it

From your Step 2 research, extract:

• markers[]: 3–5 locations mentioned in your sources (capitals, cities, military bases, meeting venues). Use real lat/lon coordinates.
• connections[]: diplomatic channels, proxy influence lines, trade routes.
• hlCodes[]: ISO alpha-2 codes of countries involved (e.g. ["IR","IQ","OM"]).

This data comes from your research. You already have it. Do not skip.

Available types

9 pre-built types (there is NO V6). DO NOT write custom D3/SVG.

V10 Big Number — VERY RARE, strict rules:

• Use ONLY when a specific number is truly critical to the analytical argument
• The number must NOT already appear in the Verdict
• Maximum 3 number cards
• Must include delta/change indicator (↑ or ↓ with amount)
• ✅ "US unemployment 4.1% (↑0.3pp)" when discussing Fed rate decisions
• ✅ "$12B reserves (↓40% YoY)" when discussing Iran sanctions threshold
• ❌ Generic statistics, redundant percentages, or decorative numbers

Placement and rules

Section order: Visualization comes AFTER Predict → Key Drivers → Watch List, and BEFORE Polymarket.

For each type you use:

1. Keep its <section> in the HTML body (after Watch List section)
2. Keep its matching <script> block at the bottom
3. Only change the data variables at the top of the script (AGENT: markers)
4. Delete ALL unused types — both HTML <section> AND <script>
5. <h2> title MUST describe what the chart shows

Label limits (auto-truncated, keep short anyway):

Step 8 — Fact Check

This report may inform six-figure financial decisions.

• Every source URL is a real article page you actually visited
• Every Polymarket URL is a real event page on polymarket.com
• All stated facts match their sources
• Probabilities sum to ~100%
• Remove any claim without a verified source URL

Step 9 — Generate HTML

Open SKILL_DIR/templates/report_template.html and read it. Then generate a new HTML file following the same structure:

1. Copy the entire <style> block verbatim — never modify CSS.
2. Copy D3 <script> blocks for the 1–2 viz types you chose. Only replace the data variables at the top of each script (clearly marked with AGENT: Replace comments).
3. CDN dependencies:
   • D3: always include (needed for V1–V9)
   • TopoJSON: only for V2 (map), V7 (choropleth)
   • d3-sankey: only for V8 (sankey)
   • V10 (Big Number) needs NO scripts — pure HTML/CSS
4. Fill in the HTML content sections following the template's structure. Section order: [A] → [B] → [E] → [F] → [G] → Visualization → [H] → [I]
5. Delete ALL unused viz types — both <section> and matching <script>.
   • No Polymarket → delete section [H]

Save as report.html in the working directory.

Step 10 — Convert to PDF

python SKILL_DIR/scripts/to_pdf.py report.html predict_report.pdf

Step 11 — Deliver

Tell the user the PDF is ready. State the verdict in one sentence: "Report generated. 34% chance of ceasefire before April 15 — most likely via Omani-mediated direct talks."

Template Quick Reference

The template contains example content for a "US-Iran Ceasefire" report. Here's the section-by-section mapping:

Domain Adaptation

Research Sources by Domain

Visualization by Domain (pick 1–2, min 1 max 2)

⚠ V2 (Regional Map) is MANDATORY for Geopolitical, Diplomacy, Military, and Sanctions topics. Do not substitute Gantt, Timeline, or any other type in place of V2. V2 must always be included; add a second chart if needed.

What NEVER Changes

1. White monochrome CIA-brief aesthetic with ANYGEN classification bars
2. 52px verdict number — the biggest element on the page
3. Verdict block: number → outcome → detail → context background → assessment
4. Probability-ranked outcomes with bars
5. Key drivers with ↑/↓ direction and causal logic
6. Watch list with forward-looking triggers and conditional impacts
7. Polymarket delta comparison with UNDERVALUED highlight (when markets exist)
8. Every URL must be a real article, never a homepage
9. 30-second scannability — no filler, no academic prose

Writing Rules