The 80/20 Rule of Prompting

80% of prompt quality comes from 4 fundamentals:

Pillar 1: Clarity

Pillar 2: Context

Without context:

"Is this transaction suspicious?"

With context:

"This merchant is a convenience store 
in Singapore, typically 50-80 txns/day 
averaging $15 SGD. Today: 340 txns 
averaging $4.50. Is this suspicious?"

Types of context: Domain · Data · Situational · Constraints

4 Types of Context

Context in Action: Merchant Review

[DOMAIN]
You are reviewing an AnyCompany Pay merchant in 
Singapore's food & beverage sector.

[DATA]
Merchant: Kopi Corner Pte Ltd (ID: MC-8842)
Monthly txn volume: 4,200 → 15,600 (6-month trend)
Avg transaction: $8.50 SGD
Chargeback rate: 0.3% → 4.1% (6-month trend)
Complaints: 12 in last 30 days (up from 2)

[SITUATIONAL]
Auto-flagged: chargeback rate exceeds 1.0% threshold.
Risk committee meets Friday.

[CONSTRAINTS]
- All amounts in SGD
- Reference AnyCompany's chargeback policy (max 1.0%)
- Use only the data provided above
- Include a GREEN/AMBER/RED risk rating

Pillar 3: Role Assignment

Pillar 4: Output Framing

Output Framing in Action

Same question: "What happened with PayLater chargebacks this quarter?"

All 4 Pillars Together

[ROLE] You are a Senior Merchant Risk Analyst at a 
Southeast Asian digital payments company.

[CONTEXT] Below is the data for a flagged merchant.
Merchant: Kopi Corner Pte Ltd (ID: MC-8842)
Sector: Food & Beverage, Singapore
Monthly txn volume: 4,200 → 15,600 (6-month trend)
Avg transaction: $8.50 SGD
Chargeback rate: 0.3% → 4.1% (6-month trend)
Complaints: 12 in last 30 days (up from 2)
Top complaint: "Item not received" (45%)

[TASK] Produce a risk assessment with a 
GREEN/AMBER/RED rating.

[OUTPUT] Format as:
1. Risk Rating (one sentence)
2. Transaction Analysis (cite numbers)
3. Chargeback Assessment (vs 1.0% threshold)
4. Recommended Actions (numbered, with owners)
Use ONLY the data provided. All amounts in SGD.

See the Difference: Merchant Review

Why Chain-of-Thought?

Financial decisions require multi-step logic. CoT makes reasoning visible and auditable.

"Can this merchant afford a 
$50K credit line?"

Answer: "Yes, they can afford it."
(No reasoning shown)

"Think through this step by step."

Revenue: $120K
- Costs: $95K = $25K free cash
- Existing debt: $15K
= Available: $10K
New repayment: $5K
DSCR = 2.0x → Affordable

CoT Techniques

Zero-Shot CoT Example

Just add "Think step by step" to any prompt — no examples needed.

Few-Shot CoT Example

Provide one example with reasoning — the AI follows the exact same pattern.

Step-Back Prompting Example

First identify principles, then apply them to the specific case.

Self-Consistency for High Stakes

Solve 3 ways, report the majority — for decisions where being wrong is costly.

See the Difference: Loan Decision

The Persona Formula

You are [TITLE] at [COMPANY TYPE] 
with [X years] of experience in [SPECIALTY].
You are known for [CHARACTERISTIC]. 
When [SITUATION], you always [BEHAVIOR].

Persona in Action: Same Merchant, Different Eyes

Data: Merchant with 271% volume growth + chargeback rate rising from 0.3% to 4.1%

Multi-Agent Framing

Get 3 perspectives in one prompt — no need to schedule 3 meetings.

Same Data, Different Audiences

Data: "PayLater default rate increased from 2.1% to 2.8% this quarter"

Why Structure Matters

Finance use cases:

• Invoice extraction → accounts payable system
• Transaction categorization → reconciliation
• Complaint classification → route to correct team
• KYC document parsing → verification forms

How to Get Structured Output

Tell the AI exactly what shape the output should take. The more specific, the more consistent.

The Best Default Format: Markdown

When you ask AI to save output as a file or produce a reusable document, Markdown wins on every dimension:

<h2 id="section-4-2">4.2 Chargeback Thresholds</h2>

## 4.2 Chargeback Thresholds

Why Markdown? The Numbers

Not just a preference — Markdown is measurably better for AI work:

The Grounding Problem

Without grounding rules, the AI mixes its training data with your documents — and you can't tell which is which.

RAG — The 4 Grounding Rules

Add these rules to any prompt where accuracy matters:

CRITICAL RULES:
1. Base your answer ONLY on the provided documents
2. After each claim, cite: [Doc Name, Section]
3. If not in documents: "Not available in 
   provided documents"
4. Do NOT use outside knowledge

DOCUMENTS:
[Doc 1: PayLater Terms v3.2]
[Doc 2: MAS Notice PSN 06]

QUESTION: What are our obligations if a customer 
misses 3 consecutive PayLater payments?

See the Difference: Policy Q&A

Meta-Prompting

Ask the AI to write prompts for you:

I need a prompt that helps our compliance team 
assess whether a new MAS regulatory update 
affects our PayLater product.

The prompt should:
- Work with any MAS circular
- Produce a structured impact assessment
- Identify operations that need to change
- Suggest a compliance timeline

Write the best possible prompt for this task.

Why Evaluate?

• Prompts degrade over time — model updates change behavior
• "It looks good" is not a metric — you need measurable quality
• Compliance requires evidence that AI outputs meet standards
• You need to compare version A vs version B objectively

Manual Evaluation: Rubrics

Process: Run same prompt 5 times → score each → average = quality score

LLM-as-Judge

Use one AI to evaluate another AI's output:

Score this merchant risk assessment:

1. All 8 sections present? (0-10)
2. Every claim cites data? (0-10)
3. Risk rating justified? (0-10)
4. Actions specific & actionable? (0-10)

Return JSON: {"completeness": X, "grounding": X, 
"consistency": X, "actionability": X, 
"total": X, "issues": ["..."]}

Run on 10 outputs → compare scores between template versions

A/B Testing Prompts

The Reality: Nobody Writes Long Prompts Every Day

You learn the techniques → build the template once → let the tools handle the rest.

Bedrock Prompt Management

Your prompt library — stored, versioned, and shared across the team.

No additional charge — you only pay for model tokens during testing.

Prompt Management: Key Features

• Prompt Templates with {{variables}} — same syntax from the exercises. Define variables with descriptions and defaults.
• Version Management — every change creates an immutable snapshot. Roll back anytime.
• Multi-Model Testing — test across Claude, Nova, Llama side-by-side. Compare quality, latency, cost.
• Up to 3 Prompt Variants — compare different versions of the same prompt to find the best performer.

Prompt Optimization (Instructor Demo)

You write a basic prompt. Bedrock rewrites it for better performance — automatically.

"Assess this merchant's risk level"

"You are a Senior Risk Analyst 
specializing in SEA digital payments.
Produce a risk assessment:
1. Rating (GREEN/AMBER/RED)
2. Transaction Pattern Analysis
3. Chargeback Assessment
4. Recommended Actions
Base analysis ONLY on provided data."

How Prompt Optimization Works

• Step 1: Submit your prompt (even a short, rough one)
• Step 2: Bedrock analyzes the prompt components
• Step 3: It rewrites with best practices — structure, constraints, model-specific formatting
• Step 4: Compare original vs optimized output side-by-side
• Step 5: Save the optimized version to your Prompt Management library

The Bottom Line

Choose Your Exercise

Open the workshop site → Prompt Engineering Exercises

7 Prompt Mistakes Everyone Makes

The Draft-Score-Revise Loop

Don't accept the first output. Build a self-improving cycle into your prompt:

Step 1 — DRAFT: Write a merchant risk summary 
  using the data provided.

Step 2 — SCORE: Rate your draft on these criteria:
  - Completeness (0-5): All required sections?
  - Grounding (0-5): Every claim cites data?
  - Actionability (0-5): Specific next steps?

Step 3 — REVISE: If total < 12, rewrite to fix 
  the lowest-scoring area. Max 2 revisions.

Output only the final version.

Break Big Tasks into Small Steps

Complex tasks fail when you ask for everything at once. Decompose instead:

"Analyze our Q2 transactions, 
identify fraud patterns, calculate 
loss exposure, compare to Q1, 
draft a board summary, and 
recommend 3 prevention measures."

Prompt 1: "Analyze Q2 transactions 
  and flag anomalies"
Prompt 2: "From these anomalies, 
  identify the top 3 fraud patterns"
Prompt 3: "Calculate loss exposure 
  for each pattern"
Prompt 4: "Draft a board summary 
  with prevention measures"

Tell the AI What NOT to Do

Positive instructions tell the AI what to include. Negative constraints prevent common failure modes:

Decision Rules: Override Subjective Judgment

Different models give different ratings for the same data. Decision rules enforce your policy:

Structure Your Prompts Like Documents

Well-organized prompts produce well-organized outputs. Use clear sections and delimiters:

### ROLE
You are a Senior Payment Operations Analyst.

### CONTEXT
<<<
[Paste your transaction data or document here]
>>>

### TASK
Analyze the data for anomalies in Thailand and Vietnam.

### OUTPUT FORMAT
- Executive summary (3 sentences)
- Anomaly table: Market | Type | Severity | Evidence
- Recommended actions (numbered, with owner)

### CONSTRAINTS
- Use ONLY the data provided above
- All amounts in SGD
- Do not exceed 400 words

Show, Don't Tell: The Power of Examples

One good example is worth 100 words of instruction:

"Categorize each transaction as 
high risk, medium risk, or low 
risk based on amount, frequency, 
and merchant type. Format as a 
table with columns for transaction 
ID, category, and reasoning."

"Categorize transactions like this:

| ID | Risk | Reason |
| T001 | HIGH | $12K single txn, 
  new merchant, no history |
| T002 | LOW | $45 recurring, 
  12-month pattern |

Now categorize these: [data]"

The 3-Round Prompt Improvement Workflow

Every production-quality prompt goes through this cycle:

Build a Team Prompt Library

Your best prompts are team assets, not personal notes. Treat them like shared templates:

Why AI "Gets Dumber" Mid-Conversation

It's not a bug — it's a context window problem. Every AI has a limited "working memory."

5 Rules for Managing Long Conversations

The Session Summary Technique

When a conversation gets too long but you can't lose the state:

The Conversation Funnel

Start broad, then narrow. Each turn builds on context — but keep it focused.

When to Start Fresh vs. Continue

The #1 Misconception: "AI Remembers Me"

It doesn't. Each session is completely isolated. Here's what AI actually sees:

"Save Your Game" — AI Memory for Long Projects

For projects spanning weeks or months, you need two files — not one giant document:

Circuit Breaker Patterns

Working Safely: Undo, Revert, Recover

What happens when the AI makes a mistake? You have safety nets at every level.

Two Controls, Two Different Jobs

These are independent settings — changing one does not affect the other.

Using Kiro for Business Users

• Vibe mode: Describe what you want → Kiro writes and runs the code
• File context: Drag CSVs, PDFs, JSON into chat
• Iterative refinement: "Make the chart bigger" / "Add a percentage column"
• New Session per task: Keep context focused

Quick Reference Card

Your Prompt Skills = Agent Design Skills

Every technique you learned today maps directly to how AI agents are built:

Preview: Templates → Skills → Automation

Tomorrow you'll turn your prompt templates into automated workflows:

The 3-Day Journey

Day 2 Outcomes

• Design prompts using the 4 pillars (Clarity, Context, Role, Output)
• Apply Chain-of-Thought and Self-Consistency for financial reasoning
• Create expert personas for different audiences
• Extract structured data and ground responses in documents
• Evaluate prompt quality with rubrics and LLM-as-Judge
• Use Bedrock tools to optimize and manage prompts at scale
• Manage long conversations and know when to start fresh
• Build reusable prompt templates — the foundation for AI agents
• Identify a workflow from your team to automate on Day 3