Analysis

Seeing the signal in the noise.

Humans are good at intuitive leaps. AI is good at processing massive volume without fatigue. Together, you solve the data overflow problem.

The Analysis Matrix

Capability	Prompt Pattern	Outcome
Summarize	"TL;DR this in 3 bullet points focusing on risks"	Speed
Extract	"Turn this messy text into a JSON object with keys: Name, Date, Action"	Structure
Classify	"Label each support ticket: Urgent, Bug, Feature Request"	Order
Pattern Match	"Find the contradiction between Document A and Document B"	Insight

Techniques

1. Few-Shot Prompting

Don't just explain; show.

"Extract company names. Input: 'I bought an iPhone from Apple yesterday.' -> Output: Apple Input: 'Microsoft released a new update.' -> Output: Microsoft Input: '[Your text here]' -> Output: "

2. Chain of Thought

Force the model to show its work. This reduces logic errors in complex analysis.

"Analyze the financial health of this company based on the text. Think step-by-step. First, identify revenue trends. Second, check debt levels. Finally, provide a verdict."

3. The Skeptic Persona

AI tends to be agreeable. Force it to be critical.

"You are a ruthless auditor. Find every logical fallacy, unproven assumption, and weak correlation in this argument."

4. Format Enforcement

Data is useless if you can't pipe it.

"Output ONLY raw CSV. No markdown, no intro text, no explanations."

The Prompt

A complete analysis prompt combining chain-of-thought, skeptic persona, and format enforcement. Paste any document, dataset, or transcript alongside it.

You are a ruthless analytical auditor. Your job is to extract signal
from noise and deliver structured, actionable insight.

DOCUMENT TO ANALYZE:
[Paste document, transcript, financial report, or dataset here]

ANALYSIS PROTOCOL:

1. EXECUTIVE SUMMARY
   - 3 bullet points, max 15 words each
   - Focus on what CHANGED, what's AT RISK, and what to DO

2. EXTRACTION
   - Pull every quantitative claim into a table:
     | Claim | Source | Verified? | Confidence |
   - Flag any number cited without a source as UNVERIFIED

3. CONTRADICTION SCAN
   - Compare all claims against each other
   - List every internal contradiction or tension
   - For each: quote both sides, explain the conflict

4. PATTERN RECOGNITION
   - What themes appear 3+ times?
   - What's conspicuously ABSENT that you'd expect?
   - What assumption does the entire document rest on?

5. SO WHAT?
   - One paragraph: if this analysis is correct, what should
     the reader DO differently tomorrow?
   - One paragraph: if this analysis is WRONG, what would
     that mean?

OUTPUT RULES:
- Tables over prose. Bullets over paragraphs.
- Every claim tagged: HIGH (data-backed), MEDIUM (reasoned),
  LOW (assumption), or UNVERIFIED
- If you catch yourself agreeing with the document, stop and
  argue the opposite position for one paragraph.

Paste into Claude, Perplexity, or NotebookLM with your source material.

Tools

Tool	Strength	Link
Perplexity	Real-time research with citations	perplexity.ai
Claude	Long context, nuanced reasoning	claude.ai
Gemini	Largest context window (whole books)	gemini.google.com
DeepSeek R1	Complex analytical reasoning	deepseek.com
NotebookLM	Source-grounded analysis	notebooklm.google

Context

• Pattern Recognition — The underlying human skill
• DeepSeek R1 — Complex analytical reasoning
• Gemini — Large context windows
• Prompts — First principles across all modalities

Mantra

Data is not information. Information is not knowledge. Knowledge is not wisdom.

Questions

What's the difference between analysis that reveals truth and analysis that confirms what you already believe?

• When does adding more data make analysis worse rather than better?
• If the AI finds a pattern, how do you distinguish signal from overfitting?
• Which of the four capabilities above (summarize, extract, classify, pattern match) do you default to — and which do you avoid?