Learn to use AI as a tool — not a search box

Welcome, Hemi.

Six weeks, one new skill each week, and a small civil-engineering project every time that you can actually check is right. By September you'll have built things, not just chatted. Work top to bottom, tick tasks as you go, and show your dad the milestone at the end of each week.

◑ ~4–5 hrs/week ◒ 100% free tools ◓ 3+ portfolio projects ◔ Local AI by the end

The golden rule

Treat AI like a brilliant but unreliable junior: fast, tireless, occasionally confidently wrong. You never sign off on a number you couldn't produce and check yourself. You own the engineering. AI does the legwork.

Your five operating principles

These apply to every task — they're how a professional will expect you to use these tools.

You are the engineer. AI is the junior.

Never accept an answer you couldn't have produced and checked yourself.

AI confidently miscalculates beam reactions and units in real tests.

Derive first, automate second.

Do the maths by hand and check it. Then let AI turn your method into a tool.

This is the workflow working engineers get hired for.

Only build things you can check.

Prefer outputs you can test against ground truth — a photo, a hand calc, raw data, a cited page.

AI dazzles on textbook problems and fails on the messy real one.

Learn in Socratic mode.

When learning, make the AI ask you questions and critique your reasoning.

Answer-first AI made students ~17% worse on a later exam.

Read every line. Show working. Cite the source.

Understand code before you run it; make AI show reasoning; ground claims in real documents.

Most AI code "runs" but only a fraction is sound — reading it is the skill.

▶ New here? Start with the 15-minute Where are you? check so the programme fits you — then Setup & tools (about an hour), then Week 1.

Start

Where are you right now?

A 15-minute calibration — not a test · retake it in September to measure how far you've come

Why this exists — so the programme fits you. No grades, nothing to revise for. Answer honestly — if you're guessing, that's useful signal too. At the end you get a personal map of what to speed through and what to slow down on. Nothing leaves this device.

Notice what this quiz is really testing

Some questions show you an AI answer that looks confident but is wrong, and ask you to catch it. That's not a trick — spotting the AI's mistake is the single most important skill in this whole programme. You own the engineering; AI does the legwork.

Answer all the questions above, then get your map.

Setup & tools

Everything here is free and runs in a browser — no software to install. Spend your first session getting these ready. Tick them off.

Pick a main chat assistant — Claude, ChatGPT or Gemini (free tier). Try the others on the same task later. Set up Google Colab — free Python in the browser (needs a Google account). Bookmark it. Create a GitHub account — where your projects will live for uni/employers to see. Open a "lab notebook" doc — one place for your prompt library + a few lines of reflection each session. Open your NotebookLM — your content library and asset factory (Dad set this up).

The big one — later

Local AI on this Mac

This MacBook Pro (M5 Max, 128 GB) can run real AI models locally — private, offline, free to use. You'll build up to it. When you're ready (around Week 5–6), install Ollama and the app's Local AI Tutor will run on your own hardware instead of the cloud. Full steps are in the README and the Asset Studio.

(Optional, later) Install Ollama and run your first local model.

Week 1

Prompting: from searching to instructing

Foundations · 3 sessions · the single highest-leverage skill

Goal — feel the difference between a search query ("what is a bending moment") and a prompt that gives AI a role, context, a task and a format.

Do

Read the prompting guides in NotebookLM; ask it to summarise the key techniques into a checklist. Learn the anatomy of a prompt — role · context · task · format · constraints. Try few-shot and chain-of-thought. Take one weak prompt and improve it five times, noting what each change did. Build a reusable study-tutor prompt that quizzes you (see the Prompt Pack). Check: verify an AI concept explanation against your NotebookLM material or a textbook. Deliverable: start your Prompt Library — 6–8 tested prompts, each with a "when to use" note. ✅ Show Dad: demo three prompts, explain why structured prompting beats search, show one thing AI got wrong.

Assets for this week

Week 2

Verify & think: catch the AI being wrong

Foundations · 2–3 sessions · the habit that keeps you safe

Goal — prove to yourself AI is an unreliable junior on maths, and use it to check and stretch your thinking, not do it for you.

Do

Take an A-level mechanics problem (forces, moments). Solve it fully by hand first. Ask AI the same problem. Compare. Where it differs, work out who's right — often it's you. Hunt a failure: a units/statics problem. Make it "show all working and check its own units." Learning mode: AI generates 3 practice variations and marks you without revealing full solutions. Deliverable: a write-up — "3 times I caught the AI being wrong." ✅ Show Dad: walk through one AI mistake you caught and how. Explain the golden rule in your words.

Assets for this week

Week 3

Data & analysis: your first code, with a partner

Building · 3 sessions · make AI write code you understand

Goal — use AI as a coding partner to load a real dataset, chart it and draw a conclusion — while reading every line.

Do

Open Colab. Pick a real dataset (rainfall, traffic, UK bridge-condition or construction data — ask AI where to find free CSVs). AI helps you load it with pandas and chart it. Rule: after every block, ask "explain this line by line." Compute one meaningful thing (trend, average, max) and write two sentences on what it means. Check: cross-check plotted values against the source for two data points. Portfolio #1: a clean Colab notebook (load → chart → finding). Save a copy to GitHub. ✅ Show Dad: run it end to end, explain any three lines he points at, show the GitHub link.

Assets for this week

Week 4

Build a tool: the beam calculator

Building · 3 sessions · derive first, automate second

Goal — turn engineering you own into a working tool. You derive the maths; AI builds the interface.

Do

By hand first: derive support reactions and max deflection for a simply-supported beam with a point load. Check it. Ask AI to build it into an interactive calculator (Streamlit app or single-page HTML): span, load, position → reactions + deflection + a diagram. Read the code. Ask it to explain anything you don't follow. Change one thing yourself. Write a short README explaining the engineering and the formula — in your words. Check: the tool must agree with your hand calc on three cases. If not, find out why — that's the lesson. Portfolio #2: a working beam calculator + README. Optionally deploy free on Streamlit. ✅ Show Dad: demo it and prove it's correct against your paper working.

Assets for this week

Week 5

RAG: make AI answer from real sources

Grounding · 2–3 sessions · answers with receipts

Goal — understand why grounding AI in documents (RAG) beats letting it answer from memory, and build an "ask-the-standard" assistant that cites its source.

Do

In NotebookLM, load a design standard, building reg or spec. Ask questions; make it quote the passage and source each time. Learn the concept: retrieve relevant text first, then answer only from that. Why does "clause + page" matter? Check: for three answers, open the real source and confirm the citation is correct. (Optional stretch) rebuild a mini "chat with a PDF" in Colab to see the mechanism. Portfolio #3: an "ask-the-standard" assistant + a one-page note on why grounding & citations matter. ✅ Show Dad: ask it a question, open the source, prove the citation is right. Explain RAG in one sentence.

Assets for this week

Week 6

Agents & capstone: put it together, then publish

Grounding · 3 sessions · your Business edge + a portfolio

Goal — see what an AI agent is (chaining steps), build a capstone that combines your skills, and publish your work.

Do

Capstone: a cost + embodied-carbon estimator — materials for your beam/a small slab, with cost and CO₂ per material, totalled. Your Business A-level edge. Make it a mini "agent": inputs → calculate quantities → look up rates → costed, carbon-rated summary. Understand each hand-off. Build in public: tidy 2–3 projects on GitHub with write-ups; record a 2-minute demo. Reflect & plan: what would make a great EPQ? How does this strengthen a uni / degree-apprenticeship application? Check: hand-check quantities and totals for one example. Every rate must have a source. ✅ Show Dad: present the capstone like a mini viva — need, method, how you verified it, AI vs you. Plan what's next.

Assets for this week

Prompt starter pack

Copy these into your Prompt Library and adapt them. The point isn't to memorise them — it's to see the shape of a good prompt, then write your own.

The Socratic Tutor

You are my A-level [Physics] tutor. I'm learning [bending moments].
Do NOT give me full solutions. Instead: ask me one question at a
time, wait for my answer, tell me if I'm right or wrong and why,
and give a small hint if I'm stuck. Start by checking what I
already understand, then build from there. Keep going until I can
solve one unaided.

The Maths Checker

Here is my solution to a mechanics problem: [paste your working].
Check it step by step. Show ALL working and explicitly check the
units at each step. If you get a different answer, show exactly
where we diverge — don't just assert yours is right. State your
final answer and your confidence.

The Code Partner

Help me build [a script that loads this CSV and plots rainfall by
month] in Python for Google Colab. Give me ONE small step at a
time. After each block, explain it line by line and tell me what
to check before running. Assume I'm a beginner and want to
understand, not just copy. Wait for me to confirm each step works.

The Grounded Answer

Answer ONLY using the document I've provided. For every claim,
quote the exact sentence and give the section/clause and page.
If the document doesn't contain the answer, say "not in the
source" — do not use outside knowledge or guess.

The Reviewer

Act as a critical senior engineer reviewing my work: [paste].
Find the three weakest points, anything I've assumed without
checking, and one thing that could be wrong or unsafe. Be
specific and blunt. Don't praise it.

Local AI Tutor

This chats with an AI model running locally on this Mac via Ollama — no cloud, private, free. It's your Socratic tutor, on your own hardware. (Until Ollama is installed, it'll show you how to set it up.)

Not connected

To switch this on (do it around Week 5–6):

Install Ollama — see the Asset Studio & Local AI tab or the app's README.md.
In Terminal run a model once, e.g. ollama run llama3.1 (downloads it, then chat/quit with /bye).
Serve this app so it can reach Ollama: from the app folder run python3 -m http.server 8000, then open http://localhost:8000.
Click Check for local AI above. If your models appear, you're running local AI. 🎉

Hi Hemi. Once your local model is connected, pick Socratic tutor and ask me to help you learn a topic — I'll question you, not just hand over answers.

Asset Studio & Local AI

Two power-ups: making your own training media with NotebookLM, and running AI locally on this Mac.

Make training media with NotebookLM

NotebookLM doesn't just answer questions about your sources — it generates media from them. Each is a real "AI beyond search" skill, and the output plugs into this app.

🎙 Audio Overview (podcast)

Two AI hosts discuss your sources. Great for revising on the go. Download the audio.

🎬 Video Overview

A narrated, slide-style walkthrough. Good for derivations. Download or link it.

📊 Infographic / slides

Turn a briefing into a visual. Screenshot to an image or export a deck.

📄 Study guide / briefing

Structured text summary — the week's reading. Save as a doc/PDF.

How to plug an asset in: save the file into app/assets/…, then add a line to app/assets/assets.json for that week. The app shows the player/image automatically. Full example in README.md. Making the materials is itself part of the training.

Going local on this Mac (M5 Max · 128 GB)

The endgame: run AI on your own hardware — private, offline, free. This machine is genuinely powerful for it.

Ollama — the easiest start. Install it, then ollama run <model> in Terminal. The Local AI Tutor connects to it automatically.
LM Studio — a friendly GUI to browse, download and chat to local models (with Apple-Silicon MLX acceleration).
Model sizes: start small (~7–8B) to feel it, then scale up — 128 GB comfortably runs 70B-class models, larger ones quantized. Browse the Ollama library / LM Studio for current models.
Then: local RAG (your own offline "ask-the-standard"), and eventually plugging into Dad's locally-hosted models and tools.

Progress

Everything is saved on this device (your account). The "Shown Dad" milestone is the weekly check-in.

0 of 0 tasks done

For Dad: the check-in guide

You don't need to know the tech. Your job is to protect the two things that make this work: that Hemi verifies, and that he understands what he built.

Five questions that do all the work

"Show me it's right." Every week has a ground-truth check — ask him to prove the output against a hand calc, the raw data, or a cited source.
"Explain this bit to me." Point at any line of code or AI answer. If he can't explain it, he's leaned on AI too hard — slow down.
"Where did the AI get it wrong?" Every week he should name something AI botched. If he never catches it, he isn't checking hard enough.
"What did you do, versus what did AI do?" Healthy: he owns the engineering and judgement; AI did the typing.
"What would you improve?" Shows he's evaluating, not just accepting.

Light touch is the point. ~30 minutes at the end of each week is plenty. Praise the catches and the checking, not the flashy output — you're rewarding judgement, the one thing AI won't commoditise.