🤖 The Modern AI Developer
The Modern AI Developer
参考斯坦福 CS146S 重制成零基础人话版。下面每篇都能直接读，想动手就进互动课。
- 1What Is an Agent?Start with an example you already get
- 2How Does an Agent 'Think'?The think–act–observe loop
- 3Hands-on: Your First Tool-Calling AgentRead a minimal snippet + judge right from wrong
- 4What Context Is: The AI's DeskWhy 'good results are a byproduct of good context'
- 5Bringing the AI Reference Material: RAG in Plain WordsLet it look things up before answering, instead of guessing from memory
- 6Let the AI Fix Itself by Looking at the ErrorFeed the error back so it can self-correct
- 7Give the AI a RoleWhy a role / system prompt helps, and how to write a good one
- 8Write a 'Manual for the AI'Intro to design docs / project specs
- 9Getting Started with AI Coding ToolsWhat exactly do Cursor / Claude Code add over a normal editor
- 10Sync vs. Async AgentsWatch it work step by step, or send it off to run to completion and report back
- 11Semi-Async Workflows: How Far to Let GoWhich tasks you can hand off, which you must watch
- 12Why Long Context DerailsFeed it too much and it forgets and drifts—how to defend against it
- 13How to Design Good Tools for an AgentSmall and clear, single responsibility, well described
- 14The CLAUDE.md Pattern: Make It Remember the Project's RulesProject-level long-term memory / convention file
- 15How Teams Share AI KnowledgeDistill personal tricks into reusable team conventions
- 16Connecting MCP: Plug a Whole Row of External Abilities into the AgentMCP in plain words: a standard plug—connect once, get a row of tools
- 17What a Modern AI Terminal Looks LikeWhy it's designed this way
- 18Run Cautious or Run AggressiveTwo configuration styles, each with its own stage
- 19When to Let Go, When to Add a CheckpointDecide with 'cost of a mistake + reversibility'
- 20One Sentence Can Trick Your AgentPrompt injection: smuggled instructions, indirect injection, and least privilege
- 21Don't Put Secrets in the Prompt, and Don't Let It RoamCredential / secret leaks and SSRF in plain words
- 22When AI Says 'Security Scan Passed,' Don't Fully Trust ItFalse negatives: passing a scan doesn't mean it's actually safe
- 23Double-Insure the AI's OutputAutomated tests + human verification as a safety net together
- 24What Code Review Is Actually ReviewingCorrectness, maintainability, team conventions—each layer harder than the last
- 25How to Review AI-Written CodeAI confidently makes things up and over-engineers—reviewing it differs from reviewing people
- 26The AI Review Capability QuadrantHand the mechanically decidable to AI, keep the context-dependent for yourself
- 27From a One-Line Idea to a Reachable Web PageBreak a fuzzy idea into clear requirements the AI can get right in one go
- 28A High-Accuracy Generation Pipeline: How to Minimize ReworkIterate in small steps, skeleton before details, verify each step before continuing
- 29UI Automation Traps and ShortcutsThe fancier it is, the easier it derails; use screenshots and automation to save effort
- 30Deploy the App to ProductionShortest path: one-click deploy on a hosting platform, domain, environment variables and secrets
- 31Monitoring & Reliability BasicsLogs, error alerts, key metrics, and what to check first when it breaks
- 32AI-Native Ops: Let the Agent Watch Production for YouThe Agent reads logs, does first-pass triage, suggests fixes; the human makes the call
- 33Evaluation & Observability: How Do You Know If Your Agent Actually WorksOffline eval + online monitoring; execution-based grading + LLM judges (which are biased)
- 34RAG & Memory: Giving Your Agent Knowledge and a MemoryParametric + non-parametric memory, reduces (not eliminates) hallucination, plus chunking / lost-in-the-middle / seven pitfalls
- 35Multi-Agent: When to Split, When Not ToWorkflow vs. agent, when it fits, plus the 15x cost and errors that compound
- 36Cost & Latency: Making Your Agent Affordable and FastPrefill / decode, output is pricier, plus caching / batching / routing / compression / speculative decoding
测一测 🧠
学完顺手检验一下，看你能认出几个。