Plan guide

What’s new in version 2

• Three buffer weeks

  Weeks 7, 13 and 17 are buffers: catch up, consolidate, rest. Nothing new is taught in them, and each one has a useful job if you’re on track. The plan runs 21 weeks: Mon 16 Nov 2026 → Sun 11 Apr 2027.

• Lighter Mon–Wed

  After a shift, the first half-hour awake is for water, daylight and food. The session is 16:15–17:00: one untimed interview problem plus revision, never new material.

• One deep project, built for a real user

  P5 (the capstone) is built for a real user: a local charity, community group, small business or team at work, with evals on their real questions. You find the user in Week 13 and build their eval set in Week 17. A fictional brief is the fallback.

• Labs vs portfolio

  P1, P3 and P4 are labs: you still build them to learn, but you pin only P2 and P5 on GitHub and talk about those in interviews.

• Sleep and overtime unchanged

  Same shifts, ≈7.5 h of sleep after every shift, a Thursday reset and a Sunday nap. Friday flex still drops in overtime weeks; anything left over goes to the next buffer week.

• Fixes from the review

  Project weeks show real weeks, the problem and drill counts agree everywhere, and leftover ‘Day 61+’ text is gone. v2.1 added the checked YouTube companion, an outdated-tutorials list and a reranker step in Week 11.

1. Step 1 · Check your Mac

   Week 1

   Apple menu → About This Mac. You should see Chip: Apple M5 and macOS Tahoe 26. Install updates and keep ≥30 GB of disk space free.

2. Step 2 · Open the Terminal

   Week 1

   ⌘ + Space, type Terminal, press Return. The text ending in % is the prompt.

   $ pwd

   $ ls

   $ mkdir -p ~/code && cd ~/code

   $ open .
3. Step 3 · Install Apple’s Command Line Tools

   Week 1

   Gives you Git and compilers. A pop-up appears: click Install and wait (10–30 min).

   $ xcode-select --install
4. Step 4 · Install Homebrew

   Week 1

   The Mac ‘app store’ for developer tools. Run the official command from brew.sh, then the ‘Next steps’ lines it prints.

   $ /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

   $ brew --version
5. Step 5 · Install uv, Git, GitHub CLI and VS Code

   Week 1

   uv manages Python versions and packages; gh logs you in to GitHub; VS Code is the editor.

   $ brew install uv git gh

   $ brew install --cask visual-studio-code
6. Step 6 · Install Python with uv

   Week 1

   Python 3.12 is a safe choice for PyTorch, MLX and most AI libraries.

   $ uv python install 3.12

   $ uv run --python 3.12 python --version
7. Step 7 · Create your project and virtual environment

   Week 1

   uv init creates pyproject.toml; uv add installs a package and records it in uv.lock.

   $ mkdir -p ~/code/ai-21-weeks && cd ~/code/ai-21-weeks

   $ uv init --python 3.12

   $ uv add requests pytest python-dotenv

   $ uv run python hello.py
8. Step 8 · Set up VS Code (AI autocomplete off for Weeks 1–4)

   Week 1

   Install the code command in PATH, the Python and Jupyter extensions, and choose the .venv interpreter. Disable any AI autocomplete until Week 5.

   $ code .
9. Step 9 · Connect Git and GitHub

   Week 1

   Choose GitHub.com → HTTPS → ‘Login with a web browser’ when gh asks.

   $ git config --global user.name "Your Name"

   $ git config --global user.email "you@example.com"

   $ gh auth login
10. Step 10 · Keep secrets out of Git

    Week 6

    Keys go in .env, .env is in .gitignore, and you commit only .env.example. If a key is ever pushed, rotate it immediately.

    $ echo 'GEMINI_API_KEY=' > .env.example

    $ echo '.env' >> .gitignore

    $ echo '.venv/' >> .gitignore
11. Step 11 · Local models: Ollama and MLX

    Week 11

    Ollama needs macOS 14+; pick a 3–4B model from ollama.com/library. mlx-lm runs and fine-tunes LLMs on Apple Silicon.

    $ ollama run <model-name-from-ollama.com/library>

    $ uv add mlx-lm

    $ uv run mlx_lm.generate --prompt "Explain RAG in one sentence"
12. Step 12 · PyTorch with the Mac GPU

    Week 15

    On Macs PyTorch uses Apple’s GPU through the MPS backend instead of CUDA.

    $ uv add torch torchvision

    $ uv run python -c "import torch; print(torch.backends.mps.is_available())"
13. Step 13 · Node.js and npm

    Week 18

    Vite (Week 19) needs Node 20.19+ or 22.12+.

    $ brew install node

    $ node --version

    $ npm --version
14. Step 14 · Claude Code, an agentic coding tool

    Week 19

    Needs Pro/Max or a Claude Console (API) account. Start it inside a project folder, never your home folder.

    $ brew install --cask claude-code

    $ claude --version

    $ cd ~/code/ai-21-weeks/projects/p5-fde-capstone && claude
15. Step 15 · Docker Desktop

    Week 20

    Download ‘Docker Desktop for Mac with Apple silicon’. Free for personal use. Quit it when you’re not using it.

    $ docker run hello-world

    $ softwareupdate --install-rosetta
16. Step 16 · Accounts to create (all free)

    Week 1

    GitHub · Google AI Studio · Claude Console · Hugging Face · Kaggle (verify phone) · Exercism · NeetCode · Render · Streamlit Community Cloud · Anthropic Academy · DeepLearning.AI · LinkedIn · Meetup.

Showing this skill in interviews without hiding gaps

• Show the log, not just the app

  Link AGENT_LOG.md in the P5 README: each entry has the task, your prompt, the agent’s plan, what you accepted, what you rejected and why, and the tests you added.

• Own every line you shipped

  If you can’t explain a piece of code in a deep-dive, it shouldn’t be in your repo. Rewrite it or remove it.

• Say who wrote what

  For example: ‘I wrote the retrieval, evals and API by hand; the agent drafted the CI workflow and the copy button, and I fixed two problems in its first attempt.’ Honest, and a strong signal.

• Tell the ‘rejected change’ story

  One concrete example of an agent suggestion you caught (it hid an error, broke a test, or added an unneeded dependency) shows the judgement employers want.

• Expect AI-free rounds

  Many coding interviews restrict AI tools, so follow each company’s stated rules. Your AI-free problems are the preparation. If AI use is allowed, narrate how you check its output.

• Don’t overclaim

  ‘I’ve used Claude Code on a handful of tasks in my capstone, with tests and review’ beats ‘I’m an expert in agentic AI’.

Free tiers and compute

• Free LLM access

  The Gemini API free tier covers Flash and Flash-Lite models; the top model is paid-only (S56). Local models run free with Ollama, MLX and LM Studio (S45, S46, S48).

• Kaggle Notebooks

  A weekly GPU quota is shown in your account (commonly reported as ≈30 h/week, with 12-hour sessions; phone verification needed) (S51).

• Google Colab

  Google says free-tier limits are not published and vary, and notebooks run for at most 12 hours (S52).

• Hugging Face ZeroGPU

  Free daily quota for small demos (S70).

• Free hosting

  Render free web service (S54); Streamlit Community Cloud (S59); GitHub for code and README demos.

• Free courses with certificates

  Anthropic Academy on Skilljar, including ‘Building with the Claude API’ and ‘Claude Code in Action’ (S58, S63); the Hugging Face Agents Course (S64).

• Free coding assistants and practice

  GitHub Copilot Free (S74) for review and explanations from Week 5. Exercism Python track (S75) and the NeetCode 150 list (S76).

• Student/startup credits

  Only if you are a student or found a company. Check the GitHub Student Developer Pack and providers’ startup programmes yourself; they aren’t assumed here.