Syllabus
The full course map — modules and units in order, plus materials and tools.
00Kickstart
42 min · 5 unitsA map for newcomers to AI coding
Goals
- Explain how roles are split in vibe coding: what you decide and check, and what the AI agent does
- Assign an AI tool to one of the four quadrants by what it does, and explain why Claude Code is an AI dev platform
- Tell the vibe coding cycle apart from the old development cycle
- Tell a primary source from packaging by three signs (licence, date of the last edit, author and participation) and explain which question this check answers and what sovereignty in learning means
- Explain in plain words token, context window, temperature, how models differ, prompt vs context, agent — and where LLMs sit in the nesting doll AI ⊃ ML ⊃ deep learning
01Introduction
30 min · 4 unitsSoftware 3.0 and why it changes everything
Goals
- Tell Software 1.0, 2.0 and 3.0 apart by how the machine's behaviour is defined in each era
- Explain why working with an agent is co-thinking and not “AI does it instead of me”, and what stays with the human
- Rewrite a command as a delegation and a spec: add context, the “why” and “what not to do”
- Decide which part of a task to hand to the agent and which to keep for yourself
- Choose which AI clone to start with, and name the steps of its build cycle
02Setup
45 min · 3 unitsSetting up your workspace for agentic coding
Goals
- Check the base environment (Node.js, Git, terminal) and use the error message to tell which tool needs to be installed
- Install Claude Code, make sure it launches, and get the terminal settings into working shape
- Explain what role AGENTS.md plays for the agent in a project folder and why CLAUDE.md holds the line @AGENTS.md
- Write an AGENTS.md the agent will follow: short, specific, with a split between “what I delegate / what I decide myself”
03Stack selection
60 min · 5 unitsWhich agentic stack fits your context: Claude Code, Sovereign, Cloud-OSS or Behind-Firewall
Goals
- Break an agentic stack down into three layers (agent/UI, routing/model gateway, model provider) and explain that each layer is swapped independently
- Choose one of the four stacks based on your constraints (budget, access, hardware, privacy, setup time) and name the trade-offs of that choice
- Explain what the LiteLLM model gateway gives you and how the fallback chain works all the way down to the local “floor”
- Tell the boss-worker scheme (Hermes) apart from a single agent that does everything itself, and determine when it's justified
- Draw up a step-by-step migration plan in which the old stack stays as a fallback until the new one has been tested on a real task
04Prompt Engineering
45 min · 5 unitsHow to write prompts like a pro
Goals
- Explain why a prompt without context gives a random result: the model fills in what’s missing with typical values
- Assemble a prompt from the parts “role, task, inputs, expected output, constraints” and assign a prompt fragment to the right part
- Match a problem in the answer with the phrase that fixes it, and tell it apart from techniques that don’t work (politeness, pressure, repeating unchanged)
- Recognize a prompting sin (context rot, lost in the middle, distractors, mixing context and prompt, aggressive tone) and choose the correction
- Choose an approach for a recurring task (meta-prompting, the reusable-assistant template) and explain why automation comes only after you can write and verify the task
05Context & Memory
35 min · 4 unitsHow agents hold information — and how to manage it
Goals
- Explain that at any moment the agent knows only what is in its “field of view”, and recognize three context problems: lost in the middle, context rot, distractors
- Separate context (stable background) from the prompt (the instruction for this request) and split them explicitly with headings or XML tags
- Choose where to store information: the project’s AGENTS.md (Claude Code imports it with the line @AGENTS.md in CLAUDE.md), the personal ~/.claude/CLAUDE.md, TODO.md, STATE.md, and tell them apart from MEMORY.md, which Claude writes itself
- Explain why a memory file must be short, and trim a bloated AGENTS.md
- Apply long-session techniques: a summary compression before a new session and parallel agents for independent subtasks
06Automation Pipeline
60 min · 5 unitsBuild your first practical workflow
Goals
- Break a recurring information flow down into trigger, action, and condition (what kicks it off, what happens, when not to run)
- Tell apart a manual, a scheduled, and an event-driven pipeline and choose the type for the task
- Build the path from manual to automatic: first run the manual pipeline as a draft and find the weak spot, then add conditions, a schedule, and notifications
- Use the automation map to judge which task can be delegated to a pipeline and which one needs a human
07Tools & Extensions
100 min · 6 unitsMCP servers, Skills, Hooks, and Superpowers
Goals
- Tell apart an MCP server, a Hook, and a Skill by purpose and choose the mechanism for the task
- Explain what MCP gives the agent, connect a server through the config, and on failure check the JSON syntax and tokens first
- Choose the hook event for the task (before or after an action, on completion) and check a hook before adding it, because it runs any command
- Turn a recurring prompt into a Skill and decide whether to keep it as personal or as project-scoped
- Assemble a minimal clone from AGENTS.md, MCP, Skill, and Hook, giving each element its role and keeping the frame, the goal, and the check for yourself
08Agent Engineering
110 min · 6 unitsFrom vibe coding to orchestrating agent systems — where the prompt stops working
Goals
- Recognize from the symptoms that a task has outgrown a single prompt, and name what that symptom needs instead of a prompt
- Break a task into steps and pick an executor for each: AI, AI with a tool, deterministic code/API, or a human
- Explain which decisions in a pipeline's architecture protect it from failure: small steps, idempotency, separating trigger and action
- Check an agent against the six blocks of production readiness, assign a problem to the right block, and name three Agent SDK settings that keep the agent within bounds: tools, permissions, and turn limit
- Write an agent specification before the code: honestly mark the gaps and tell a spec from a charter
09AI Notebook
38 min · 5 unitsSource-grounded learning: a notebook that answers only from your sources — and how to check where it lies
Goals
- Tell a source-grounded notebook's answer from an answer of a regular chat and of a hybrid, and explain why a citation makes an answer checkable
- Predict what the notebook won't see in a source: what is shown but not said, and an empty source
- Check a notebook answer against the inspection checklist and choose an action for a warning signal
- Choose an output format for the consumption scenario and decline a format that has no scenario
- Assemble a notebook for your own task: a topic with a question, source selection, a first question about disagreements, inspection depth by how much the answer matters
10Model Trainingadvanced
42 min · 5 unitsAn advanced module after the agentic practice: how LLMs are built and trained, why data is the main work, when to fine-tune and when a prompt and context are enough — and your own fine-tune of a small model on a free GPU
Goals
- Explain how tokens, embeddings, attention and pretraining on next-token prediction add up to an LLM — and why the same design produces confident mistakes
- Lay out your task along the machine learning cycle and pick an algorithm class by what you need to get — including the conclusion "no LLM needed here"
- Decide whether fine-tuning is needed or the task is closed by a prompt, context/RAG and agents; choose between full fine-tuning and LoRA and pick a small model
- Plan an honest evaluation: set the holdout aside before training, recognize overfitting and forgetting, tell a benchmark apart from a check on your own task
- Fine-tune a small model on a free GPU or your own hardware and record a "before / after / prompt without fine-tuning" comparison against criteria set in advance
11Second Brain
50 min · 5 unitsA second brain that stays with you: memory in your own files, a channel to reach your agent, skills that travel from agent to agent, and security rules for everything that comes from outside
Goals
- Explain what a second brain that stays with you is made of (memory in files, an agent, a channel), and ask your archive so that every claim in the answer has a file address
- Describe how a message travels from a messenger to the agent and back, connect a channel following the official documentation, and keep only yourself on the allowed-senders list
- Explain what actually travels when "one agent teaches another" — skill files and templates, not model weights — and hand over a skill through a folder or a repository
- Recognise untrusted input — messages, emails, other people's skills — and apply the rules: read a skill before installing, do not act on requests from incoming messages, keep keys out of messages, use a separate token, grant least privilege
- Build a minimal second brain — memory, one channel, one skill from someone else — and confirm it with three checks and a report in my-experiments
Course Materials
Templates
From the course
Going deeper
- Microsoft agents course — a Russian guideOur guide to the 18 lessons: what to read after which module, and what to skip
- AI Agents for Beginners (original) ↗18 lessons from Microsoft on agent patterns, MCP, memory and security. MIT, updated regularly
- Hugging Face Agents Course ↗Hugging Face course on agents: how they work, agent frameworks in practice, a final project. Free, Apache-2.0
- DeepLearning.AI short courses ↗Hour-long courses on agents, RAG and orchestration. Free, but a closed platform: sign-up required, no open licence
Foundations: models and data
- Generative AI for Beginners ↗How large models work inside: tokens, embeddings, fine-tuning. Microsoft, MIT
- Machine Learning for Beginners ↗Classical ML from before the LLM era — the ground everything stands on. Microsoft, MIT
- Data Science for Beginners ↗Working with data: collection, cleaning, visualisation. Microsoft, MIT
- Hugging Face LLM Course ↗Transformers and the Hugging Face ecosystem from the inside — theory for the Model Training module. Apache-2.0
- smol-course: fine-tuning small models ↗A hands-on Hugging Face course: SFT, LoRA, evaluation on small models — the basis of the Model Training practice. Apache-2.0
- The Smol Training Playbook ↗Hugging Face on how LLMs are actually trained and fine-tuned: data, pre-training, post-training. Reading for the Model Training module
- LLM Evaluation Guidebook ↗Hugging Face on how benchmarks work and how to evaluate a model for your own task, including your own "vibe tests"
Process for bigger projects
The agent harness
- OpenAI: Harness engineering ↗How an OpenAI team built a product in five months where Codex writes all the code and humans build the harness: docs, checks, linters
- LangChain: Improving Deep Agents with harness engineering ↗Same model, different harness: +13.7 points on Terminal Bench 2.0, a loop detector, a reasoning-depth "sandwich"
- ETH Zurich: are AGENTS.md files helpful for coding agents? ↗A 2026 study: context files don't raise success on average and add cost; only non-standard rules help. CC BY 4.0