# 03 · Facilitator Runbook — Minute-by-Minute > Stage-manager's cue sheet. Designed to sit on your second monitor (or printed) during the live session. Read top-to-bottom on Monday morning; glance at the timing column once you're live. > > Times in this document are **session-relative** (0:00 = the moment you start broadcasting), not wall-clock. The wall-clock for the 18 May 2026 session is 19:00 Bucharest, but treat that as a separate sheet on your monitor. > > Cross-references: slide numbers and lab details match `02_slide_deck.md`. Strategic background lives in `01_course_design.md`. This doc assumes you've read both. --- ## Pre-flight checklist ### T–60 minutes · Sit down, no interruptions - [ ] **Read this entire runbook once.** Yes, even if you read it last night. - [ ] **Apply the P0 slide patches** if you haven't already (full list in `02_slide_deck.md` appendix). Budget 25 minutes. If you're behind on time, prioritise slides 12, 14, 15. - [ ] **Test impress.js deck** in the browser you'll share. Use Chrome — Safari sometimes misrenders transforms. Confirm right-arrow navigation works and the gradient backgrounds render. - [ ] **Plan B for the deck.** Export a flat PDF as backup. If impress.js breaks mid-session, swap to the PDF and keep going. - [ ] **Open `companion.html` in a separate tab.** Confirm Lab 1, Lab 2, Lab 3 panels are reachable. Confirm the commitment field saves to localStorage. - [ ] **Open two LLM tabs as the demo tools.** Claude.ai and ChatGPT, signed in. If one is down mid-session, switch without explanation. - [ ] **Open a private text file** for scratch — chat moves fast and you'll want a place to jot questions you can't answer immediately. ### T–30 minutes · Tools and infrastructure - [ ] **Zoom (or your platform) configured:** - Cloud recording ON. - Local recording ON (belt and suspenders). - Chat enabled for all participants. - Co-host added if you have one. - Breakout rooms disabled (we're going single-broadcast for Lab 3 unless you've decided otherwise — see `01_course_design.md` §16). - [ ] **Camera and mic check.** Two-minute monologue into the recording, play it back, confirm levels. - [ ] **Close every other app.** Notifications off. Slack quit. Email quit. Phone face down. - [ ] **Water on your desk.** You'll talk for 2 hours and 20 minutes. ### T–10 minutes · Open the room - [ ] **Open the broadcast room.** Let early arrivals in with a casual mic-off welcome message in chat. - [ ] **Pin the companion link to the chat** so anyone joining late can grab it. - [ ] **Share screen showing slide 1.** Music optional — if you do music, fade it at 0:00. - [ ] **Take a breath.** Three deep ones. You've got this. ### T–1 minute · Pre-roll - [ ] **Recording confirmed ON.** Verbally: *"Recording is on."* If a co-host is present, they confirm. - [ ] **Final chat post:** *"We start in 60 seconds. Companion link is pinned — open it in a second tab now."* --- ## Live schedule The time column shows session-relative minutes. The cue column gives you the line that begins each block — read it as written, or paraphrase if it lands more naturally for you. ### Welcome block — 0:00 to 0:10 (10 min) | Time | Slide | Move | Cue | |---|---|---|---| | 0:00 | 1 | Open | *"Welcome. I'm Florin, and for the next two and a half hours..."* | | 0:01 | 1 | The bar | *"By the end of today, you've used AI on a real task in your work."* | | 0:02 | 1 | Housekeeping | Recording, mics, chat, companion link. | | 0:03 | 2 | Show agenda | *"Three arcs. One promise. Three labs."* | | 0:04 | 2 | Wave check | Ask for a 👋 in chat. Confirms audio and chat both work. | | 0:05 | 2 | Read room | If anyone says "no audio" / "can't see slide" — fix now, not later. | | 0:07 | — | Ground rules | Questions go in chat. I batch them at module breaks. Mics off unless speaking. | | 0:09 | — | Bridge | *"Let's start with how this stuff actually works."* Move to module 1. | **Key facilitator moves in this block:** - Project warmth in the first 30 seconds. No apologies. No "hopefully." - The 👋 emoji check at 0:04 is non-negotiable — it surfaces audio issues while you can still fix them. - If anyone joins after 0:05, don't backtrack. Acknowledge in chat: *"Recording is on; we'll send it tomorrow."* --- ### Module 1 — 0:10 to 0:45 (35 min) | Time | Slide | Move | Note | |---|---|---|---| | 0:10 | 3 | SKIP | Don't show the module header. Just talk. | | 0:10 | 4 | Family tree | 45 seconds. Don't elaborate. | | 0:11 | 5 | History | 75 seconds. The 2017 transformer paper is the key beat. | | 0:12 | 6 | Neural nets | 2 minutes. The 86B neurons + brain analogy. | | 0:14 | 7 | How AI learns | 2.5 min. **The $100M training cost is a strong moment — pause on it.** | | 0:17 | 8 | What is an LLM | 2.5 min. Emphasise "autocomplete at extraordinary scale." | | 0:19 | 9 | Tokens | 2.5 min. The "strawberry r's" example lands well. | | 0:22 | 10 | Prediction | 2.5 min. "Probabilistic — same question can give different answers." | | 0:25 | 11 | Transformers | 45 sec. One sentence on attention. Skip the with/without example. | | 0:26 | 12 | Models | 3 min. **Use the rewritten card content from `02_slide_deck.md`.** | | 0:29 | 13 | Hallucinations | 3.5 min. **Slow down. This is the most important slide of the day.** | | 0:33 | 13 | 3-step protocol | Repeat it twice. Cite. Check exists. Check says. | | 0:35 | 14 | Context window | 2.5 min. The "no longer the bottleneck" reframe. | | 0:37 | 15 | Knowledge cutoff | 2 min. "6–12 month sabbatical" analogy. | | 0:39 | 16 | Recap | 45 sec. Quick checklist, then break. | | 0:40 | — | **BREAK** | *"Ten minutes. I'll be back at the top of the hour."* | **Branch logic:** - **If running ahead at 0:25** (≥3 min ahead): on slide 11, take 2 minutes for the with/without attention example after all. Group C will appreciate it. - **If running behind at 0:25** (≥3 min behind): skip the slide 11 elaboration entirely. On slide 12, name only OpenAI, Anthropic, Google — skip Meta unless asked. - **If running >5 min behind at 0:35:** cut slide 16 entirely. Trigger break immediately after slide 15. **Questions in chat:** scan during slides 9, 10, 14, 16. Batch and answer at 0:39 if any are quick. Defer anything technical to module 2 break or end-of-session. --- ### Break — 0:45 to 0:55 (10 min) | Time | Move | |---|---| | 0:45 | Companion timer triggered. "Capture one thing you learned" field highlighted. | | 0:46 | You: stand up, stretch, drink water, glance at chat. | | 0:50 | Skim chat for any "tools not working" reports. Reply briefly. | | 0:53 | Post in chat: *"Back in 2 minutes. Lab 1 starts at the top of the hour."* | | 0:55 | Mic on. *"Welcome back. Now the part you came for."* | **Do not start early.** Some people use the full 10 minutes. Starting at 0:53 means restarting people who walked away. Hold the line. --- ### Module 2 — 0:55 to 2:05 (70 min) | Time | Slide | Move | Note | |---|---|---|---| | 0:55 | 17 | SKIP | Don't show the module 2 header. | | 0:55 | NEW-A | Day 1 starters | 2.5 min. **The bridge from theory to practice.** | | 0:58 | 18 | Tools landscape | 3 min. Update with the patched content. | | 1:01 | 19 | Chat anatomy | 1.5 min. Fast. | | 1:02 | 20 | What is a prompt | 1.5 min. The freelancer analogy. | | 1:04 | 21 | Specific | 2.5 min. Two-column comparison. | | 1:06 | 22 | **RTFC** | 3.5 min. **Drill it. Highest-stakes slide.** | | 1:10 | — | **LAB 1** | 8 min. See lab section below. | | 1:18 | 23 | Examples | 2 min. Show-don't-tell. | | 1:20 | 24 | Format | 1.5 min. Quick. | | 1:21 | 25 | Iterate | 2.5 min. Three follow-up turns example. | | 1:24 | — | **LAB 2** | 12 min. See lab section below. | | 1:36 | 26–29 | Use cases | 4 min. **Use the merged single slide.** | | 1:40 | 30 | What AI is NOT | 3 min. The four cautions. | | 1:43 | 31 | Privacy | 2.5 min. Update with Team/Enterprise carve-out. | | 1:46 | 32 | Human in loop | 2.5 min. **The skill shift moment.** | | 1:48 | NEW-B | MCP & agents | 2.5 min. Frontier glimpse. | | 1:51 | — | **LAB 3** | 14 min in lab + 6 min for sharing/closing the lab. | | 2:05 | — | Bridge to close | *"Coming home stretch."* | **Branch logic for module 2:** - **If LAB 1 runs over (>8 min):** absorb time from slides 23–24. Compress to a single 2-minute beat covering both. RTFC and iteration are higher-value than examples and format. - **If LAB 2 runs over (>12 min):** trim the use-cases tour from 4 to 2 minutes. Drop the brainstorming and coding quadrants verbally; people will see them on the slide. - **If you're under time at 1:48:** spend the buffer on NEW-B (MCP/agents). Group C will love it and it sets up the advanced course pitch. **Questions in chat:** scan during transitions, especially between slides 25 and Lab 2 (people often have follow-up questions about iteration). Batch and answer briefly at 1:23. Anything that takes >60 seconds defers to end of session. --- ### Close + 30-day plan — 2:05 to 2:20 (15 min) | Time | Slide | Move | Note | |---|---|---|---| | 2:05 | 33 | Daily workflow | 2.5 min. Week 1 / week 2 framing. | | 2:08 | 34 | Resources | 1.5 min. Don't read the URLs aloud — they're in the workbook. | | 2:09 | NEW-C | **Commitment** | 5 min total: 1 min framing, 2 min silence while they type, 2 min readout. | | 2:14 | — | Q&A | 4 min. Take 3–4 from chat. Anything else: "in the recording or follow-up email." | | 2:18 | 35 | Thank you | 1.5 min. **End on warmth, not slides.** | | 2:20 | — | Stop recording | Confirm in chat: *"Recording stops here. Thank you, all."* | **The commitment moment is sacred.** Don't talk over the 2 minutes of silence. Don't fill it with riffing. Let them type. Look at your screen, not the camera. If you're tempted to say something, count to ten instead. **Q&A protocol:** - Pick questions from chat that benefit the whole cohort, not narrow personal ones. - Anything personal/narrow: *"Good question — happy to address in the T+7 check-in email."* - Hard technical questions: *"That's an advanced-course topic — let me note it for the resource list."* - Hostile or out-of-scope questions: don't engage live. Acknowledge politely, move on. --- ### Buffer — 2:20 to 2:30 (10 min) This block is for you, not for the participants. **Do not extend the session into it.** - Stop the recording at 2:20 exactly. - Stay on for 2–3 minutes for anyone who wants to wave goodbye. - Then close the room at 2:23–2:25. - Use 2:25 to 2:30 for your own decompression before you do anything else. If you've genuinely run short (rare — happens if labs end early), bring forward NEW-B's MCP/agents content as an expanded segment. Or take one final group-C technical question. Or — best option — just end early. Ending a 2.5-hour live session 5–10 minutes early is a gift, not a failure. --- ## Lab transitions — the three fragile moments The labs are where the session can break. Everything before each lab is preparation; everything after each lab is reinforcement. Follow these scripts closely the first time you run the course; loosen as you build experience. ### Lab 1 transition · "Vague → Specific" **Start cue (≈45 sec, read tightly):** > "Eight minutes. Switch tabs to your LLM — ChatGPT, Claude, doesn't matter which. Take the prompt 'Write something about marketing' and rewrite it using RTFC. Role, Task, Format, Constraints. Run it in your LLM. Paste your *rewritten prompt* — not the model's response, just your prompt — into our chat when you're done. I'll pick two or three to read aloud. Camera off if you want. Eight minutes starts now." **Companion action:** click "Reveal Lab 1" — the lab panel becomes visible. **Mid-lab moves:** - **+3:00** — Post in chat: *"Halfway. If you're stuck on the role, pick any role and start there — you can refine later."* - **+6:00** — Post: *"Two minutes. Drop your rewritten prompt in chat."* - **+8:00** — Mic on. Read 2–3 prompts aloud, with attribution to the participant. Say one specific thing you like about each. End with: *"Notice — they all have a role, a task, a format, a constraint. None of them is genius. That's the point."* **Common failure modes:** - *No one is pasting in chat at +6:00.* Add 90 seconds. Post: *"Even a rough version — paste what you have."* - *Someone pastes a model response instead of a prompt.* Don't call it out publicly. Skip past it to the next one. - *Someone goes off-topic with a question during the lab.* Reply briefly in chat — don't break the lab energy. **Exit condition:** at least 3 prompts in the chat. Read 2–3 aloud. Move to slide 23. --- ### Lab 2 transition · "Three-turn improvement" **Start cue (≈45 sec):** > "Twelve minutes for Lab 2. Take your prompt from Lab 1, or start fresh with a new task — your choice. Run the first turn. Then iterate three times. Turn 2: change the format or length. Turn 3: change the tone or audience. Turn 4: ask 'what's missing?' Pay attention to how the output evolves. When you're done, drop one sentence in chat: what changed turn to turn. Twelve minutes starts now." **Companion action:** click "Reveal Lab 2." **Mid-lab moves:** - **+5:00** — Post: *"Halfway. Make sure you've done at least one 'what's missing' turn — that's where the gold is."* - **+9:00** — Post: *"Three minutes. Wrap your loop."* - **+11:00** — Post: *"One minute. One sentence in chat about what changed."* - **+12:00** — Mic on. Read 2–3 chat sentences aloud. Reinforce: *"Notice — none of you got the best output on turn one. That's the lesson."* **Common failure modes:** - *Many participants only do one turn.* Push back live: *"If you stopped after turn one, you got the worst version of what's possible. Two more turns is the difference."* - *Someone says the model got worse with iteration.* Acknowledge as real — *"Yes, sometimes iteration goes sideways. That's why you save the version you like and start a new turn from there."* **Exit condition:** at least 4 sentences in chat. Read 2–3 aloud. Move to use-cases tour. --- ### Lab 3 transition · "Your real task" — the lab the course pivots on This is the lab. Get this one right and the cohort leaves changed. Get it wrong and the workshop is just another workshop. **Pre-lab framing (≈75 sec):** > "OK. The lab the whole course is built around. Twenty minutes — the longest stretch of the day. You take your real task — the one you brought with you from the pre-work, or pick one of the five starters on the slide — and you run it. > > Three things before you start. One: *anonymise.* If your task involves customer names, company financials, anything sensitive — swap the names. Don't paste real PII or NDA-protected data. Two: don't aim for the perfect output. Aim for *better than you'd have done without AI in twenty minutes.* That's the bar. Three: if you get stuck, drop a question in chat — I'll triage as I see them. > > Twenty minutes. When you have something usable, drop one sentence in chat: 'My task was X. The biggest surprise was Y.' Cameras and mics off, work mode on. Starting now." **Companion action:** click "Reveal Lab 3." **Mid-lab moves — this is where you earn your facilitation:** - **+3:00** — Scan chat for stuck-people signals: *"I don't know what to ask"*, *"my model just gave me X"*, *"I'm not sure if this is good"*. Post 1–2 quick triage tips publicly. Examples: - *"If you don't know where to start: pick starter #1 — rewrite a difficult email you've actually been putting off."* - *"If your model is giving generic output: add a Role line — 'Act as my friend who's blunt and direct.'"* - **+10:00** — Post: *"Halfway. If you're still on turn one, push to turn two — that's where the lift happens."* - **+15:00** — Post: *"Five minutes. Aim to land your output, not to polish it forever."* - **+18:00** — Post: *"Two minutes. Drop your task-and-surprise sentence in chat."* - **+20:00** — Mic on. Read 3–4 chat sentences aloud, by name. Affirm specifically — not just "great work" but *"that's a real workflow you can use Tuesday morning."* **The closing line for Lab 3 — say it slowly and mean it:** > "That's the win. That's why you're here. Most of you just used AI to do something you'd have done another way an hour ago, and it worked. The skill is now real. Keep doing reps." **Common failure modes:** - *Some participants don't post their task/surprise.* That's normal. Don't push. Read what you have. - *Someone is visibly frustrated in chat.* Send a direct private message during a break: *"Saw your message — want a quick follow-up after the session?"* Don't engage publicly mid-flow. - *The first 2 readouts are weak/generic.* Pick from later in the chat — there are almost always strong ones at +18:00 from the people who used the full time. **Exit condition:** at least 5 sentences in chat. Read 3–4 aloud. Move to slide 33. --- ## Q&A protocol — the rules of the chat Why a protocol matters: in a 2.5-hour live session with 25 strangers, free-form Q&A destroys momentum. A clear protocol respects everyone's time. **The rules (state at 0:07, enforce throughout):** 1. **Questions go in chat, not unmute.** Unless invited. 2. **I batch and answer at module breaks.** Don't expect immediate answers. 3. **Mark urgency.** Add `[URGENT]` if the session is broken for you (audio dropped, can't see slides). That bypasses the batching. 4. **Personal/narrow questions:** I'll address in the T+7 follow-up email. **Three pre-prepared deflection lines for when you need them:** - *"Good technical question — that's an advanced-course topic. I'll add it to the resource list."* - *"That's a great use case for the Coaching tier 1:1 — happy to go deep there."* - *"Excellent question. Hold it for me — I'll address in the recap email."* These let you respect the question without derailing the schedule. **Scan rhythm during the live session:** - Skim chat during transitions only, not while presenting. - Co-host (if present) flags `[URGENT]` markers and the top 3 most-upvoted questions. - Solo-host fallback: budget 30 seconds at each module boundary to skim. --- ## Technical contingency What to do when things break. Sorted by likelihood. ### Audio drops (your end) - Reconnect on phone or laptop within 60 seconds. - If you can't reconnect: chat *"audio issue — back in 90 sec — keep your companion open."* - Don't panic. Audience patience is generous if you communicate. ### Screen share fails - Stop and restart screen share. Usually fixes it in 10 seconds. - If impress.js misrenders: stop sharing, switch to the PDF export, restart sharing. - Time cost: ≤45 seconds if you've practised. ### LLM tool you're demoing is down - Switch demos to your second tool tab. Don't announce the switch — just keep going. - If both are down: pivot to a screenshot from your prep folder. - *"As you can see in this earlier example..."* covers the swap. ### Participant audio chaos (someone forgot to mute) - Mute them centrally from the host panel. Don't call them out. - Move on without comment. ### Recording fails - You have two recording sources (Zoom cloud + local). One usually survives. - Confirm at 0:00 and at 1:25 (just after the break). - If both have failed mid-session: don't tell the participants live. Address it in the T+1 email — *"recording had a technical issue, here's a summary instead."* ### A participant becomes hostile or disruptive - Don't engage live. Move them to "attendee" role (no chat, no unmute) from host panel. - Acknowledge to the room only if it was visible: *"Apologies for the interruption — let's continue."* - Follow up privately after the session if appropriate. --- ## Voice and tone — small calibration notes A few specific facilitation moves that pay off: **Use participant names.** When you read someone's chat aloud, attribute it: *"Maria says... that's exactly right."* People light up when their name is said in a workshop of 25. Costs you nothing. **Pause after big lines.** "AI is a first draft, not a final answer" — pause two seconds. Let it land. The instinct is to keep talking; resist. **Don't apologise for the format.** No *"sorry we have to move fast"* or *"I wish we had more time for X."* You designed the time budget. Own it. **Don't read the slide.** Participants can read. Your job is to add the layer that isn't on the slide — examples, stakes, the why. **Energy peaks at three moments:** the bar at 0:01, the RTFC drill at 1:06, the Lab 3 closing at 2:11. Be sharper at those three than anywhere else. Mid-slides between them can be more conversational. **The hardest moment is 1:35 (after Lab 2).** Energy dips. People are 90 minutes in. The use-cases tour is what carries you through. Make it punchy, fast, vivid. Specific examples beat abstract categories. --- ## Post-session checklist ### T+0 to T+15 minutes · While it's still fresh - [ ] Stop recording. Confirm files are saving. - [ ] Save the chat transcript (Zoom → "Save chat" or screenshot). - [ ] Note any commitments you made live ("I'll cover that in the follow-up", "let me send you that link"). - [ ] Walk away from the screen for 5 minutes. You earned it. ### T+1 hour · The hot wash-up - [ ] In a private note: what worked, what dragged, what to change for next cohort. - [ ] Save chat transcript with timestamps. - [ ] Note any participant who said something quotable — they're testimonial candidates if they consent. - [ ] Note any participant who seemed stuck or frustrated — they're worth a personal email. ### T+24 hours · The follow-up email - [ ] Send to all participants: - Recording link - Slide deck PDF - Workbook (final version with their commitment captured if you've got companion data) - Companion link (so they can revisit it from any device, even though their commitment is local-only) - For Coaching tier: Calendly link to book the 1:1 - One paragraph summary of the session — what they accomplished - The T+7 promise: *"I'll check in next Monday."* ### T+7 days · The check-in - [ ] One question to each participant: *"Did you use AI for something this week? Reply with one sentence."* - [ ] Replies become: - Testimonial material (with consent) - Signal of who's stuck (they didn't reply, or replied "no") - Input for the T+21 prompt-of-the-week pack ### T+30 days · The cohort retro - [ ] Send invite to a free 30-min group call. - [ ] Soft pitch for the advanced course at the end. - [ ] No pressure. People who want the next step take it. --- ## The single most important sentence in this runbook Every cue, every patch, every facilitation move serves one outcome: > **The participant uses AI on a real task during Lab 3, and again in the seven days after.** If a decision in the moment helps that outcome, take it. If it doesn't, don't. Everything else is decoration. Good luck on Monday.