Placeholder cover — replace with real title art (swap this content for  once the image exists).

By the end, everyone has built something they can actually use — a piece of content, a communication plan, a draft three-year business plan — using their own practice as the input.

Notice how far this has moved since the last time you looked. A jarring open before we step back — this is where things actually are, not where you last checked. Flag: the "doubling every 3-4 months" figure is the longer-run historical average (mid-2024 to Mar 2026); the Mar-to-Sep 2026 projection alone is a 6-month doubling, not 3-4 — worth deciding which framing to use before presenting.

Two words worth keeping: training and inference. Training is finished and expensive; inference is what you're doing all day and costs fractions of a penny. It explains the knowledge cutoff, and it explains why the thing that felt like it "got to know you" yesterday starts blank today.

The single most useful thing to understand today. It explains why context has to be supplied rather than assumed, why the same question gets a worse answer in a tired thread than a fresh one, and why we're going to keep our own copy of everything that matters. Comes back after the prompting exercises.

IMAGE PLACEHOLDER — drop your own example in here. Something that shows the model reinterpreting a real thing rather than producing a photorealistic fake; the point lands better with a personal image than a stock one. Add as: 

The practical version: "here is everything we give a new patient — where does it repeat itself, and where does it contradict?" That question was impossible to ask a machine two years ago.

The through-line for the whole day. If you take one thing from this section: the constraint is no longer access or money, it's knowing what to ask for. That's what the next hour is about. Note: no figure on this slide on purpose — if you want a number here it needs sourcing before the day.

Each line holds capability fixed and asks what it costs. GPT-3.5-level general knowledge fell about 9x a year; GPT-4o-level science reasoning about 900x. Note the log scale — every gridline is a tenfold drop. The blue line is the one to point at: the hardest capability is also the fastest-falling, and it starts where the others end. That's the pattern behind "wait six months and it's free". Caveat if asked: the series runs to early 2025, so it understates where we are now.

The one piece of planning that pays off later: a table that spends all morning on Gemini can hit its limit exactly when Part 4 starts. Caps aren't published as numbers and change often, so don't quote figures — "roughly, and watch for the warning". If someone's work Google account has Gemini switched off by their admin, a personal account works.

Everyone runs this at their table. The output will be competent, readable, and completely useless — it could have been written for any practice, in any country, for any patient. That's the point. Don't explain why yet; let them see it.

The instinct is to close the window and conclude AI isn't useful. The productive move is to say so and keep going. Note what comes back: it will name the gaps for you — location, client group, purpose, voice. That list is the raw material for the next prompt.

Run it and compare against the first output side by side. The difference isn't length or effort: this one says who's asking, who it's for, how long, in what voice, and what the piece is meant to do. Ask the room which parts of their own version they had to think hardest about.

Do this now, at the tables, before going on — it takes two minutes and everything after today depends on it. This is the answer to the context-window slide: the tool won't hold your context for you, so you hold it yourself.

This is the rescue move for anyone who's stuck. It converts a vague request into a well-specified one without them having to write the specification themselves. Works on any task, not just content. Worth saying: if the questions it asks are bad, that tells you something too.

The line is identifiability, not clinical content — you can discuss the case, you can't hand over the person. Everything typed into an AI tool goes to someone else's server, and several free tiers train on inputs by default. Worth checking yours.

The framework arrives last, on purpose — they've already built one of these by hand over the past four prompts, so this names what they did rather than teaching it cold. Not every prompt needs all five. Context and constraints are the two most often missing.

Usually it returns something close to role/context/task/constraints/output — which is a useful check on the framework rather than a replacement for it. Knowing the structure is what lets you tell whether the prompt it wrote is any good. This pattern comes back in Part 4.

The file is the instruction because it's the only thing that works identically across four tools in one room, moves work between them, and survives hitting a cap. Projects are a convenience on top, not a replacement — Claude free caps you at five and without retrieval, and none of it transfers if you switch tools. Anyone who hits a limit today: new conversation, paste the file, carry on.

The recap, not the introduction — they've just done all five of these by hand. Read it as a checklist against what they built rather than as new material, then open the floor.

Let this run. The questions people ask here tell you what to emphasise in Parts 2 to 4.

If AI-generated content can out-engage professional marketers on cold clicks, what could it do for content your patients already want to open? Today's session teaches the accessible version of this.

The prompting habits you just practised in Part 1 are what make this work.

This is theirs, not the model's — pens down on the laptops. The five-name list is deliberately cheap; the strategic thinking is in which one they pick and how much of step 3 they can fill in unaided. Two minutes on the list, three on the chosen group. Then ask a few tables which group they picked and why. The gaps they hit in step 3 are the argument for the next slide.

Now the model earns its place. The point of the four parts is that the first two come from them and can't be guessed; the model is filling gaps, not inventing a customer. Expect two or three rounds — if a table gets a usable persona from one prompt, they probably accepted something generic. If they picked a growth group they don't yet serve, the clarifying questions will expose what they don't know about them, which is the more valuable output. Everything after this — the newsletter, the website review, the communication plan, the business plan — is tested against this one person.

The one they picked, worked up properly. Notice what makes it usable: a named job she's hiring you for, the objections in her own words with the answer she'd accept, and where in her path to booking you currently lose her. A persona that stops at "58, back pain, self-pay" tells you nothing you can act on. Don't expect this in one prompt — this is after two or three rounds of pushing back.

Written for the persona on the previous slide, not for "patients" in general — the 3am question is hers, the compost is her garden, the Market Rasen hours are her clinic. That's the whole argument for building personas first. The photo slots are marked rather than generated: the images stay your job.

"Professional but conversational" is a register, not a voice — it's what makes AI content sound like AI content. This is the prompt that makes a draft sound like your practice. The six rules are reusable: save them and paste them into every future content prompt.

Where the time saving actually lives. One piece of thinking, five places it can go — and it feeds straight into Part 3, where you decide which of those channels you'll actually use.

The wrap, not another exercise — they've built all of this live. Worth naming the pattern out loud: the persona is the unit of decision, and everything downstream gets tested against it. Hold the room until the file is updated. Part 3 opens by asking them to paste the persona and the content back into a fresh conversation, and it doesn't work if the file is empty.

The kind of website and competitor research you'd usually put off — increasingly, you don't have to do it yourself at all.

Builds directly on Part 2 — the content you just built needs somewhere to go, and someone to trust it.

Walk through this live with screenshots rather than pasted text: it shows that AI can read a page as a page — layout, hierarchy, what catches the eye first — not just its words. Six slides follow: the site, the findings, the proposed order, where to start, and two mockups. Demonstration only; they run the text version themselves afterwards.

A real Lincolnshire practice, seven clinics, thirty years trading. Nothing wrong with it that isn't wrong with most practice websites — which is the point. Ask the room what they'd book, and how.

Eleven problems, ordered by impact on bookings, each paired with a fix. Note the ordering: it isn't a list of everything wrong, it's a list of what to do first. That came from asking for it that way.

Same content, resequenced around the decision a visitor is actually making. No new copy, no new photography, no design spend — just a different running order.

Three columns: this week with no design needed, one build sprint, and the bigger bets. The left-hand column is the one that matters — those are changes a practice owner can make themselves on a Friday afternoon.

Compare against the first slide in this series. It names the service, the county, and the number of clinics; it gives a phone number you can press and a booking button; and it says how soon you'd be seen. All of it was already true — none of it was on the page.

Nobody wakes up wanting physiotherapy — they wake up with a bad shoulder. Four entry points instead of one generic booking button, and the insurer strip answers the question everyone asks on the phone.

The reframe is what makes this work on a free tier — "review as a prospective client in pain" gets you something usable, where "do an SEO audit" gets you a generic checklist. The ranking and the developer flag turn the output into a to-do list rather than a wall of advice.

Trust cuts both ways. Dismissing AI advice outright makes patients rate you as less engaged, not more credible.

When people bring a problem to AI rather than a task, health is the single biggest thing they bring — 27%, ahead of career, relationships, and money. Read the right-hand column aloud: interpreting test results, chronic conditions, injuries and treatment. That is your consultation, happening without you, at a scale no practice can see. The previous slide says patients arrive having had that conversation; this one says how many. It's also the argument for the policy on the next slide — if this is where they go first, being explicit about how you use AI is table stakes.

Two minutes at tables, then take three responses. Don't resolve it — the next slide turns it into something they can use.

The reflection you just did, made into an artefact. Everyone in the practice will face this conversation, not just the clinicians — which is why the second half matters more than the first.

One page, printable, split by who needs what: reception uses the first two steps, clinicians use all four. The two lists at the bottom are the part to read aloud — the right-hand column is what most of us say without noticing, and every line in it ends the conversation. Worth asking the room which of those they've used this month.

Nothing new to invent here — every input is already in their file. If a table is stuck, it's because the file is thin, not because the task is hard.

One asset, six touchpoints, sequenced so each one feeds the next — and every row has an owner and a measure, which is what separates a plan from a wish list. Point at the sequencing logic: the service page goes first because everything else links to it, and the highest-effort item goes last because by then it recycles material that already exists. The three panels along the bottom are the Part 3 spine in one view — the website change, the staff line, and helping patients ask AI better questions.

The wrap. Three named strands going in — website, client-facing, staff-facing — one plan coming out. The file should now have six sections; Part 4 opens by asking them to paste all six into a fresh conversation, so this is the last chance to fill the gaps. Worth saying: nothing in that plan came from a template — every line of it traces back to the persona they built before the break.

Remember Part 1's 16-hour unattended AI run? A ~32-hour version, unattended and multi-step, is roughly what tonight's report is a taste of. What you'll do today is the accessible version of this.

Builds directly on Parts 2 and 3 — your personas, your content, your communication channels are today's raw material. Here, it's built with AI rather than about AI. Continuity, if the room raises it: some of them will have seen Celia Champion's IPPTA session on one- and three-year business goals (Painless Practice). Same territory, different tool — say it verbally rather than putting it on screen. This was Tim's steer on scope (planning call, 24 Aug 2026), not delegate-facing content.

Worth seeing a finished one before starting your own. If you run this on Gemini free rather than a paid tool, it matches exactly what the room is about to do — and you can say so.

The prompt, and the plan it proposes before it starts. Worth pausing on the plan card — it shows the model deciding what to go and read.

It works through the plan itself, ticking off steps. This runs for up to twenty minutes.

Executive summary, with local figures and inline citations back to the sources.

Policy context — ICB priorities, PCN and FCP routes.

Strategic objectives, each tied back to something in the evidence above.

Year-by-year actions with KPIs against each.

Three scenarios rather than one forecast.

Risks, impact, mitigation.

Gantt timeline and a referral-pathway flowchart it generated for itself.

The same Deep Research output, condensed to something you'd actually put in front of a bank or a business partner. Worth saying out loud: the numbers here are modelled, not audited — the value is the structure and the sequencing, and every figure in it needs your judgement before anyone else sees it.

This is the pay-off for saving Gemini capacity this morning. The cap means most people get one shot, which is the right pressure: it makes them build the prompt properly rather than firing off three mediocre ones. Deep Research may also be unavailable at peak times for free users — if it is, the interview route is the fallback, not a failure.

For everyone not on Gemini, and for anyone whose Deep Research is capped or unavailable. The one-at-a-time constraint is the whole trick: it turns a single prompt into a structured interview, which is most of what Deep Research is doing anyway. Without it you get ten questions in a wall of text and nobody answers any of them.

Deliberately unstructured. Reports are running in the background; this is the time to catch what didn't surface earlier.

The most important prompt of the day. AI writes fluently whether or not it has anything behind it, and a plan that reads well is the easiest thing in the world to over-trust. This is the habit to leave with: whatever it produces, make it attack its own work before you act on it.

Turns four sessions of activity into one page you can actually act on.

Placeholder back cover — replace with real closing art (swap this content for  once the image exists).