🗺 Skills Atlas — Tra theo kỹ năng
Skilljar tổ chức theo course (info-delivery container). Trang này tổ chức theo skill (capability bạn đạt được). Một skill có thể được dạy qua 2-5 bài rải rác nhiều course — tra ở đây để biết exact lessons.
Skilljar organizes by course. This page organizes by skill — a single capability often spans 2-5 lessons across multiple courses. Look up by skill, not by course.
📋 Cấu trúc
Mỗi skill có:
- Mức độ (foundational / intermediate / advanced)
- Định nghĩa ngắn
- Lesson dạy
- Skill phụ thuộc (cần học trước)
Tag con [F] foundational, [I] intermediate, [A] advanced.
1. Prompt Engineering
1.1 Basic prompt structure [F]
Viết được prompt rõ ràng, có context, có yêu cầu output cụ thể.
1.2 System prompts [I]
Phân biệt system vs user prompt, dùng system để define persona/constraints.
1.3 Few-shot prompting [I]
Cung cấp 2-5 ví dụ in-context để định hướng output format.
1.4 Step-by-step reasoning [I]
Yêu cầu Claude "think step by step", improve reasoning quality.
- 06.27 Structure with XML tags — wrap reasoning trong tags
- 17 — AI Capabilities and Limitations — hiểu next-token & working memory
1.5 Extended thinking [A]
Bật mode "think before output" cho task phức tạp.
Phụ thuộc: 1.1 → 1.2 → (1.3, 1.4) → 1.5
2. Context Management
2.1 Context window basics [F]
Hiểu 200K token = ~500 trang, 1M token cho Pro.
2.2 Projects (Claude.ai) [F]
Persistent context per project, upload files, instructions.
2.3 CLAUDE.md (Claude Code) [I]
Per-repo memory, hooks behaviour, sub-agents.
2.4 Prompt caching (API) [A]
Cache 90% off cho repeated context, giảm cost lớn.
Phụ thuộc: 2.1 → (2.2, 2.3) → 2.4
3. Tool Use
3.1 Defining a tool [I]
JSON schema cho function, Claude quyết định khi gọi.
3.2 Multiple tools [I]
Cho Claude gọi nhiều tool cùng lúc / chuỗi nhiều tool.
3.3 Fine-grained tool use [A]
Streaming partial JSON khi tool đang được build.
3.4 Built-in tools (text edit, web search) [A]
Built-in tools không cần tự define schema.
Phụ thuộc: 1.1 + 2.4 → 3.1 → 3.2 → (3.3, 3.4)
4. RAG / Retrieval
4.1 RAG concept [I]
Khi nào dùng RAG vs context dài.
4.2 Embeddings & vector search [I]
Voyage embeddings, cosine similarity, top-k.
4.3 Hybrid retrieval (BM25 + vector) [A]
BM25 lexical + vector embeddings cho recall cao.
4.4 Reranking & contextual retrieval [A]
Reranker cho precision sau retrieval — chỉ có ở Bedrock/Vertex track.
Phụ thuộc: 4.1 → 4.2 → 4.3 → 4.4
5. Agents & Multi-step
5.1 Single-step Claude [F]
Một prompt → một response.
5.2 Cowork task loop [I]
Claude.ai's Cowork mode — multi-step với scheduled tasks.
5.3 Agentic loop (API) [A]
while-loop tool calls cho đến terminal state.
- 06.37 Multi-turn conversations with tools + 06.38 Implementing multiple turns
- 12.84 Agents and workflows — workflow patterns (chỉ ở Vertex track)
5.4 Subagent architecture [A]
Lead agent dispatch sub-agents để giảm context size.
5.5 Skills (reusable workflows) [A]
Đóng gói workflow thành skill markdown share-able.
Phụ thuộc: 5.1 → 5.2 → 3.* → 5.3 → (5.4, 5.5)
6. MCP (Model Context Protocol)
6.1 MCP concept [I]
Standardize tool/resource expose từ external system.
6.2 MCP server [I]
Build server expose tool/resource cho Claude consume.
6.3 MCP client [I]
Build client integrate vào Claude/Cursor/IDE khác.
6.4 Transport (stdio / HTTP) [A]
Khi nào dùng stdio vs HTTP, security implications.
6.5 Sampling [A]
MCP server "ask" Claude qua client, agent-of-agents.
6.6 Roots & notifications [A]
File system access, real-time updates.
Phụ thuộc: 3.1 + 4.1 → 6.1 → (6.2, 6.3) → (6.4, 6.5, 6.6)
7. Evaluation
7.1 Eval concept [I]
Test set, scoring, regression detection.
7.2 Eval datasets & runs [A]
Build dataset, run eval pipeline.
7.3 Grading: model-based & code-based [A]
Use Claude (or rubric code) to score Claude output.
- 06.20 Model-based grading — LLM-as-judge
- 06.21 Code-based grading — exact match / rules
Phụ thuộc: 1.1 → 7.1 → 7.2 → 7.3
8. Safety & Discernment
8.1 4D framework [F]
Delegation, Description, Discernment, Diligence.
8.2 AI capabilities & failure modes [F]
Khi nào AI bịa, làm sao detect — gốc rễ ở next-token prediction & working memory.
8.3 System prompts as guardrails [I]
User input có thể override — system prompt set tone & constraints.
8.4 Steerability & character [F]
Vì sao Claude từ chối câu hỏi nhất định.
Phụ thuộc: 8.1 → (8.2, 8.4) → 8.3
9. Cloud Deployment
9.1 Bedrock fundamentals [I]
AWS-flavored Claude API.
9.2 Vertex AI fundamentals [I]
GCP-flavored Claude API.
9.3 Compliance (HIPAA, FedRAMP) [A]
Chỉ Bedrock và Vertex hỗ trợ sẵn — xem trang gốc Anthropic về compliance + AWS BAA / GCP DPA.
Phụ thuộc: 1.* + 3.* → (9.1 hoặc 9.2) → 9.3
10. Education-specific
10.1 4D in pedagogy [I]
Áp dụng 4D vào lesson design.
10.2 AI as learning partner [F]
Sinh viên dùng AI mà không gian lận.
10.3 Assessing & assignment design [I]
Đánh giá 4Ds, design assignment có AI.
11. Agent Engineering (Harness)
11.1 Harness architecture [I]
Hiểu 5 hệ thống con: instructions, state, verification, scope, tools/environment.
11.2 Progressive instruction disclosure [I]
Tổ chức CLAUDE.md/AGENTS.md thành 3 tầng: entry → topic → inline.
11.3 Multi-session state persistence [I]
PROGRESS.md, DECISIONS.md, git checkpoints cho continuity xuyên phiên.
11.4 Feature list as executable constraint [A]
JSON-based feature tracking with verification commands per feature.
11.5 Three-layer verification [A]
Syntax → runtime → system-level validation pipeline.
11.6 Agent observability [A]
Runtime + process observability, evaluator rubrics, sprint contracts.
Phụ thuộc: 2.3 (CLAUDE.md) + 5.4 (Subagents) → 11.1 → 11.2 → 11.3 → (11.4, 11.5) → 11.6
🌐 Skill heatmap (v1.1 coverage)
| Domain | F | I | A | Total skills |
|---|---|---|---|---|
| Prompt Engineering | 1 | 3 | 1 | 5 |
| Context Management | 2 | 1 | 1 | 4 |
| Tool Use | 0 | 2 | 2 | 4 |
| RAG | 0 | 2 | 2 | 4 |
| Agents | 1 | 1 | 3 | 5 |
| MCP | 0 | 3 | 3 | 6 |
| Evaluation | 0 | 1 | 2 | 3 |
| Safety | 3 | 1 | 0 | 4 |
| Cloud | 0 | 2 | 1 | 3 |
| Education | 1 | 2 | 0 | 3 |
| Agent Engineering | 0 | 3 | 3 | 6 |
| Total | 8 | 21 | 18 | 47 |
47 skills across 11 domains. Foundational skills (8) là entry barrier thấp — bắt đầu từ đây nếu bạn newbie.
🔁 Liên kết
- By use-case — "Tôi muốn làm X"
- Time-boxed paths — chọn skill học theo thời gian
- Common mistakes — skill nào dễ sai
- Dependency map — course-level dependency