Chapter 2: AI Harness · From Principles to Product Implementation | xueai.miyang.cn

Chapter 2

AI
Harness

From context management to Agent deployment, from cost control to security defense — this is the practical AI Harness handbook for AI Product Managers.

Sections

1

Context Engineering

2

Prompt Engineering

3

Security: Prompt Attack & Defense

4

Agent Engineering

5

Cost Optimization

Instructor Tip

Click any slide on the right to jump directly, or click "Start Teaching" to play slides in order.
Each slide's content unfolds naturally as you scroll; continue with "Next Lesson" at the bottom.

Context Engineering

Context Window: AI's Working Memory

Visualize window composition, drag to simulate overflow, compare capacities across major models

Context Overflow: Three Handling Strategies

Direct truncation / Summary compression / Selective retention — visualize the trade-offs of each strategy

Prompt Engineering

Why LLMs Choose Markdown

Plain-text model + formatting needs = MD as the top choice; step-by-step reasoning why HTML/Word/LaTeX don't work

Markdown Syntax and Rendering Pipeline

Quick syntax reference + live editor + marked.js / react-markdown / markdown-it rendering options

You Say It, It Becomes It

Switch between five roles in real time, control output format, core principles of System Prompt

Advanced Prompt Techniques

Few-Shot / CoT / constraints / task decomposition — good-vs-bad comparison interactive demo

Output Format Trade-offs

Plain text / JSON / Markdown / YAML / XML — scenario-fit comparison, Token cost vs. parsing complexity trade-offs

Streaming Output and Format Compatibility

JSON requires full text to parse / MD renders word by word / XML captures and renders tags on the fly — live demo of the differences

Security Defense

Prompt Injection: Why Do Attacks Succeed?

SQL injection analogy → Message List fundamentals → lack of parameterization → overview of 5 attack types

Security Step-through

Prompt Injection: 12 Attack Case Studies

Privilege escalation / role-play / Few-Shot / structural injection / metaphor disguise — vulnerable vs. defended version for each

Security Interactive

Prompt Defense: TiedStory Three-Layer Interception in Practice

Input layer regex → prompt layer constraints → output layer leak detection → frontend + secondary review, full attack-chain simulation

Hands-on Interactive

Agent Engineering

Agent: AI That Gets Things Done

Four core capabilities: Plan / Tool / Memory / Act — click to see real-world examples

The Secret of Tool Calling

Model outputs JSON → framework parses and executes → result injected back — four-step process visualization

ReAct in Practice: The Real Loop an Agent Goes Through

Step-by-step Thought / Action / Observation demo — 3 failures + 1 success — the complete Agent struggle loop

Skill: Help the Agent Take the Right Path

Skill = process instructions + tool-calling guidance; good loop vs. bad loop analogy with SKILL.md file structure

Scaffolding Engineering: From Prototype to Product

Simulate Agent booking flights and hotels — full comparison with vs. without scaffolding, 5 key capabilities explained

Cost Optimization

Why Does Multi-Turn Dialogue Get More Expensive?

Visualize cumulative Token costs — drag turns to see how costs grow exponentially

KV Cache: Trading Space for Time (and Money)

Analogy-based understanding + savings calculator — explicit cache comparison: Anthropic vs. Qwen in practice

Explicit Caching: Anthropic vs. Qwen in Practice

cache_control syntax, cache hit detection, price discount comparison — real savings demonstrated

Dynamic Timestamps: The Most Expensive System Prompt Design

Bad design vs. good design — switch between three time-handling approaches

Holistic Cost Optimization: Save Money at the System Level

5-layer optimization strategy, cost breakdown visualization, system design that saves 70–90%

Image Tokens: Pixels That Burn Money

Image billing formula, scaling mechanism, resolution pitfalls, task-based tiered strategy

Match Resolution to Task: Fine-Grained Image Cost Control

High / medium / low resolution strategies — Token cost comparison across scenarios with selection advice

Syntax-Level Optimization: Prompts Written for Machines

YAML vs JSON, CSV vs arrays, compressed JSON output — save 10–30% on formatting Tokens

Prompt Engineering

Semantic-Level Optimization: Don't Use Context as a Trash Can

Dynamic Few-Shot, LLMLingua-2 compression, drag-to-sort attention — boost information density

Output Layer + KV Cache Advanced

Negative constraints, Diff polishing, stop sequences; KV Cache tool traps and sliding window issues

Model Selection: Capability vs. Cost Trade-offs and Decision Framework

Capability/cost matrix for major models, selection decision tree, model-matching strategy by scenario

The Greatest Simplicity: Holding to First Principles

The essence of AI Harness / what to build vs. what to skip / Harness that will be obsoleted / the ultimate question

Closing Reflection

Chapter 2 Summary (Part 1) · Prompt Engineering + Agent Design

Context overflow strategies / 6 elements of Prompt / truth about tool calling / Skill + scaffolding

Chapter Summary

Chapter 2 Summary (Part 2) · Cost Optimization + Full AI PM Perspective

Five-layer cost system / KV Cache principles / Image Tokens / complete course capability checklist

Chapter Summary