# Understanding AI & LLMs — Your Workbook **Workshop:** Monday 18 May 2026 · 2.5 hours · live online **Facilitator:** Florin Bădiță · ai-courses.badita.org --- ## Welcome This workbook is your companion through the workshop and the four weeks that follow. It has three jobs. Before the session, it sets you up so the live time is spent on the work, not on definitions. During the session, it gives you a place to capture what matters and worksheets for the three labs. After the session, it's a reference you can revisit — the frameworks on one page, the 30-day plan, the resource list. There's also an interactive companion at the URL in the welcome email. The companion handles the live stuff — the lab prompts revealed on cue, your written commitment, a system-prompt builder. Both work together. Keep the companion open in a browser tab during the session; keep this workbook open as your reference and reflection space. You don't need to print anything. You don't need to fill in every blank. Use the parts that help you and skip the rest. --- ## Before the session — your pre-work Total time: about 20 minutes. None of this is graded, none of it is checked. It exists because doing these three small things will roughly double how much you get out of the live session. ### Pre-work item 1 · Sign up for one LLM tool Pick one of these three and create an account. Free tier is fine for everything we'll do. - claude.ai — Claude (Anthropic) - chat.openai.com — ChatGPT (OpenAI) - gemini.google.com — Gemini (Google) If you already use one, stick with it. If you've never used any, **pick Claude** — it tends to be the most forgiving for beginners and has the fewest aggressive upsells in the free interface. Confirm you can send one message and receive a reply. That's the bar. ☐ Done — I can send and receive messages in my chosen tool. ### Pre-work item 2 · Send this baseline prompt Open your tool. Paste this prompt exactly as written: > *In one paragraph, what's the difference between AI, machine learning, and a large language model? Explain it the way you would to a smart friend who doesn't work in tech.* Read the response. Notice your reaction — does it make sense? Does it raise questions? Is it the explanation you would have given? **Save the response.** Copy it into the box below or screenshot it somewhere you can find it later. At the end of the session, we'll come back to this — you'll explain the same concept yourself, and the comparison is one of the most satisfying moments of the workshop. **Your baseline response (paste here):** ``` [ paste the LLM's response here ] ``` ☐ Done — I sent the prompt and saved the response. ### Pre-work item 3 · Bring one real task This is the most important piece. Pick one real task from your actual work that you'd like to try using AI on during the session. We'll work on it together in Lab 3. Three criteria for a good choice: - It's something you'd do anyway in the next week or two. - It takes 10 to 30 minutes when you do it the normal way. - It has a draft-then-polish shape — there's some kind of output that exists at the end. Boring is good. Boring is great. The most banal task is the most valuable one because you have a clear baseline to compare against. **Examples that work well:** - Rewrite a difficult email you've been putting off - Summarise a long document you have open - Draft a meeting agenda from a list of topics - Generate interview questions for a candidate or guest - Prepare for a difficult conversation by getting AI to role-play it - Outline a presentation or proposal - Translate something into a language you don't speak fluently - Brainstorm names, taglines, or angles for something - Critique a draft you wrote yourself **Write your task here (one sentence is enough):** ``` The task I'm bringing: ________________________________________________ ________________________________________________________________________ ``` If you can't think of one before the session, that's fine. We'll have starter tasks ready in the companion. ☐ Done — I have a real task in mind for Lab 3. ### Optional · One short read If you have an extra 15 minutes and want to walk in warm: - Ethan Mollick's *Co-Intelligence* — first chapter, available as a sample on most ebook stores - learnprompting.org — the "Basics" section, about 10 minutes - Anthropic's *Building Effective Agents* — readable even for non-technical audiences Skip this if you're tight on time. It's not assumed knowledge. --- ## During the session — what to expect You don't need to take detailed notes. The session is recorded and you'll get the slide deck and the recording the next day. Use this workbook to capture the things that matter to you specifically — the moments where you think "that applies to my work." **The structure of the day:** 1. **Welcome and the one promise** (10 min) 2. **Module 1 — How AI actually works** (35 min) 3. **Break** (10 min) 4. **Module 2 — How to use AI** (70 min, with three labs interleaved) 5. **Your 30-day plan** (15 min) There are three live labs: - **Lab 1 — Vague → Specific** (8 min). You rewrite a vague prompt using the RTFC framework. - **Lab 2 — Three-turn improvement** (12 min). You iterate three follow-up turns on a response. - **Lab 3 — Your real task** (20 min). You use AI to produce a real output for the task you brought. Lab 3 is the moment the whole workshop is built around. By the end of those 20 minutes, you'll have something usable that you produced in the session itself. Open the companion in a second tab when the session starts. The lab prompts live there. --- ## Lab worksheets You can do the labs entirely in the companion if you prefer. These worksheet pages are here as an alternative or as a place to reflect afterwards. ### Lab 1 worksheet · Vague → Specific (8 minutes) **Starting prompt (given):** *"Write something about marketing."* **Your rewritten prompt using RTFC:** ``` Role: _________________________________________________________ Task: _________________________________________________________ _________________________________________________________ Format: _________________________________________________________ Constraints: _________________________________________________________ _________________________________________________________ ``` **Your full prompt, assembled:** ``` ``` **The model's response — what was good, what was off:** ``` ``` **Reflection:** Which of the four RTFC elements was hardest to write? That's the one to practise first. ``` ``` --- ### Lab 2 worksheet · Three-turn improvement (12 minutes) **Your starting prompt (use Lab 1's or start fresh):** ``` ``` **Turn 1 — model's first response. Note what you like and what you'd change:** ``` What I'd change after turn 1: ``` **Turn 2 — change the format or length. Your follow-up:** ``` ``` **Turn 3 — change the tone or audience. Your follow-up:** ``` ``` **Turn 4 — ask "what's missing?". Note what the model surfaced:** ``` ``` **Reflection:** Which turn produced the biggest improvement? Most people find it's turn 4. Was it for you? ``` ``` --- ### Lab 3 worksheet · Your real task (20 minutes) This is the lab. By the end of it, you'll have a real usable output. **The task:** ``` ``` **Anonymisation check** — before you start, change any: - Real customer or client names → "Customer X" - Real company financials → ranges or placeholders - Anything covered by NDA or that contains PII → genericise ☐ Anonymised — I'm safe to paste this into a consumer LLM. **My prompt (use RTFC):** ``` ``` **Iterations** — at least two follow-up turns: ``` Turn 2 (what I changed): Turn 3 (what I changed): ``` **The final usable output** — paste or summarise: ``` ``` **One sentence about what surprised me:** ``` ``` --- ## The frameworks — your reference card The two pages every participant has tape-marked in their workbook a month later. ### RTFC — the prompt framework When you write a prompt for anything that matters, run through these four in your head: **R — Role** Who do you want the model to be? *"Act as a senior product manager at a B2B fintech."* — *"Act as a skeptical investor reading a pitch deck."* — *"Act as a copy editor with a brutal eye."* The role primes the model's tone and the knowledge it draws on. **T — Task** What specifically do you want it to do? Not "help me with my deck" — that's not a task. *"Write a one-page brief outlining the problem, the solution, the key metrics, and the risks."* That's a task. **F — Format** What shape is the output? Bullets, table, numbered list, JSON, email with subject line, executive summary in three sentences. Say it explicitly. Models default to whatever shape was most common in their training data — rarely what you want. **C — Constraints** What are the boundaries? Word count. Tone. Audience. Things to avoid. *"Under 400 words. No jargon. Audience is our non-technical CEO. Don't use the word synergy."* **Worked example:** > *Act as a senior product manager at a B2B fintech startup. Write a one-page product brief for a new expense-tracking feature, with sections: Problem, Solution, Key Metrics, and Risks. Under 400 words. Avoid technical jargon. Audience is our non-technical CEO.* Every clause is doing one of the four jobs. Copy the structure, change the nouns. ### The 3-step verify protocol When the model gives you a specific fact — a citation, a statistic, a date, a quote, a price — anything you'd need to defend in a meeting: 1. **Ask the model to cite its source.** 2. **Check the source exists.** Search for it. Open it. 3. **Check the source says what the model claims.** Step 3 catches what step 2 misses. Models will sometimes cite real papers that don't say what they claim those papers say. This takes 90 seconds and it's the single biggest protection against the most common AI failure mode. ### The five prompting principles In short form: 1. **Be specific.** Vague in, vague out. Always state the task, audience, format, constraints. 2. **Add context.** Use RTFC — role, task, format, constraints. 3. **Give examples.** Even one or two examples dramatically improves output. Especially for tone and style. 4. **Ask for a format.** Bullets, table, JSON, email. Two seconds in the prompt saves five minutes of reformatting. 5. **Iterate.** Best outputs come on turn 2 or 3. Useful follow-ups: *"Make it shorter."* *"More formal."* *"Add an example."* *"What's missing?"* ### What AI is reliably good at vs. not | Reliably good at | Be cautious about | |---|---| | Drafting first versions of writing | Citing specific facts, dates, statistics | | Summarising long documents | Recent events (check the cutoff date) | | Generating options to choose from | Medical, legal, financial advice | | Explaining concepts at any level | Counting letters or simple arithmetic | | Reformatting and restructuring text | Anything that needs domain expertise to verify | | Translating between languages | Tasks where being wrong has real consequences | | Brainstorming and devil's advocate | Anything covered by NDA or with PII | ### The human-in-the-loop rule A human reviews, validates, and takes responsibility for any AI output before it has real-world effect. The AI drafts. You decide. Your value used to be *"can you write a good email?"* AI can. Your value is now *"can you judge whether this email is right?"* That judgment — your domain expertise, your context, your taste — is the irreplaceable part. Lean into it. --- ## Privacy — what never to paste These rules apply on the free and personal-paid tiers of every consumer LLM. Enterprise/Team plans usually have contractual protections, but assume defaults unless you've checked. **Never paste:** - Customer or patient PII (names, emails, IDs, records) - Passwords, API keys, secrets of any kind - Company financial data not yet public - Medical records - Proprietary source code your employer would consider confidential - Anything covered by an NDA **Safer options:** - Anonymise before pasting. *"John Smith at ACME Corp"* becomes *"Customer X at Company Y."* You get the same AI help without the risk. - Disable chat history in your tool's settings (most tools allow this). - For your organisation: ChatGPT Team, ChatGPT Enterprise, Claude Team, and Claude Enterprise plans all contractually exclude your data from training. - For maximum privacy: run a local model with Ollama or LM Studio. Slower and weaker than frontier models, but nothing leaves your machine. --- ## Your 30-day plan The workshop ends today. The course ends 30 days from now, if anything actually happens in those 30 days. Here's the structure. ### Week 1 · Reps, not theory Every day this week, use AI for at least one of these three things: ☐ **Draft one email.** Pick a difficult one — declining something, asking for something, addressing a complaint. Paste the situation; let AI draft; you edit. ☐ **Summarise one document.** Anything long that you'd normally skim. Paste it; ask for "the three key things and one action I should take." ☐ **Explain one thing you've been avoiding.** Some concept you keep meaning to understand. Ask AI to explain it in three different ways at increasing levels of detail. The goal this week is not to use AI for hard things. The goal is to make using AI a reflex. Day-by-day check-in (just tick if you did it): | Mon | Tue | Wed | Thu | Fri | Sat | Sun | |---|---|---|---|---|---|---| | ☐ | ☐ | ☐ | ☐ | ☐ | ☐ | ☐ | ### Weeks 2 to 4 · Expand Once the reflex is in place, broaden: - **Use it as a sounding board before decisions.** Describe the decision; ask AI to argue both sides; notice what you hadn't considered. - **Let it structure your meeting agendas and notes.** Paste raw thoughts; ask for a clean agenda or a structured summary. - **Ask it to critique your work before you share it.** Paste a draft; ask "what's the strongest pushback someone could make against this?" - **Use the 'what's missing' question.** On everything. It's the highest-leverage follow-up. ### The seven-day check-in In about seven days, you'll get an email from me asking one question: > *Did you use AI for something this week? Reply with one sentence.* Even if the answer is "barely" — reply. The replies help me know who needs a nudge, and they often become testimonials for future cohorts (with your permission). It also means I'm a real person on the other end, not a drip campaign. ### Common stalls and what to do **"I keep forgetting to try it."** Put a sticky note on your laptop for one week. "Try AI first." **"My outputs feel mediocre."** Re-read the RTFC card. Specifically — are you giving a role and constraints, or just a task? **"I tried it and it was wrong."** Good. That's training your trust calibration. Use the 3-step verify protocol from now on. **"I don't know what to use it for."** Re-read the five starter tasks at the top of this workbook. Pick the most boring one. Boring is the point. --- ## Resources Curated tightly. Everything here earns its place. ### Tools to try (all have free tiers) - **claude.ai** — Claude, by Anthropic. My default for most writing and reasoning. - **chat.openai.com** — ChatGPT, by OpenAI. The most widely used; great defaults. - **gemini.google.com** — Gemini, by Google. Best if you live in Google Workspace. - **perplexity.ai** — AI + real-time web search with citations. Best for research and current events. - **ollama.ai** — Run models locally on your own machine. For when privacy matters most. ### To go deeper (books) - *Co-Intelligence* by Ethan Mollick — the best non-technical book on living and working with AI in 2025–26. - *The Coming Wave* by Mustafa Suleyman — the wider lens, where this is all going. - *AI Engineering* by Chip Huyen — if you want to build with AI, not just use it. ### To stay current (low-noise) - **Stratechery** by Ben Thompson — paid, but the AI essays are sharp. - **Import AI** by Jack Clark — weekly, technical but accessible. - **The Rundown AI** — daily, lighter, good for keeping a finger on the pulse. Check in once a month at most. AI moves fast but most of the news doesn't actually change how you should work. ### To go technical (videos and courses) - **3Blue1Brown's neural networks series** on YouTube — the best visual explanation of how this works, free. - **Andrej Karpathy's "Let's build GPT"** on YouTube — three hours, technical, brilliant. - **fast.ai** — free, hands-on, opinionated, for those who code. ### When you're ready for more The advanced course — *Advanced ChatGPT & LLMs* — goes deep on attention mechanisms, RLHF, RAG, embeddings, chain-of-thought reasoning, agents, MCP, and building production systems. It assumes everything in today's workshop as background. Don't take the advanced course before you've used today's material for at least a month. Reps first, theory second. ai-courses.badita.org --- ## A note for Coaching-tier participants If you're on the Coaching tier ($129), you'll see a calendar link in your follow-up email tomorrow. Book your 30-minute 1:1 in the next two weeks while everything is fresh. To prepare for that call: - Use the system-prompt builder in the companion to draft a starter system prompt for your role. - Pick three prompts from your real work that you want feedback on. Paste them and the model's responses into a doc. I'll review them and we'll talk through where they can be sharper. - Bring one question that didn't get answered in the session. You'll leave the call with a personalised system prompt (`CLAUDE.md` style) tailored to your work, plus written feedback on your three prompts. That's the artifact you keep. --- ## Your facilitator Florin Bădiță. Civic activist, builder, TEDx speaker, Davos House of Collaboration tech director. I've spent the last two years using AI tooling aggressively across multiple production systems — and the principles in this workshop are the ones that survived contact with real work, not the ones that sound good in talks. If you want to keep in touch: - ai-courses.badita.org — courses and writing - The T+7 email — I really do read every reply See you on the other side.