4 weeks. 10 lessons.
4 things you will ship.
A practical, hands-on course for business professionals who want to build real AI tools — without a technical background. Every week ends with something deployed and working in your business.
What students leave with
By the end of this course, students will
Know what large language models can and cannot do, and how to apply them effectively in a business context
Use the 5-Part Framework to produce consistent, high-quality AI outputs for any repeating work task
Create multi-step AI processes in Flowise or LangFlow without writing any code
Convert their workflow into a shareable application their team can use from a browser link
Connect AI to existing business tools using Zapier or n8n so processes run 24/7 without manual input
Calculate time saved, cost reduction, and business impact — and document it for stakeholders
Who this course is designed for
Week-by-week breakdown
- What AI actually is — and isn't
- How ChatGPT, Claude, and Gemini differ
- When AI works and when it fails
- The 5-Part Prompt Framework (Role, Context, Task, Constraints, Output)
- How to write prompts that work first time
- The refinement loop — how to improve outputs
- Prompt Cards — documenting for reuse
- Compare ChatGPT vs Claude on the same question
- Build a 5-Part prompt for a real work task
- Test with 3 inputs — easy, real, edge case
- Refine and document in a Prompt Card
- Calculate time saved per week
- ChatGPT (free tier)
- Claude by Anthropic (free tier)
- Google Gemini (optional)
- Why single prompts aren't enough for complex tasks
- Workflow thinking: Input → Process → Output
- When to chain AI steps and when to stop
- Introduction to Flowise (chatflow builder)
- Introduction to LangFlow (pipeline builder)
- How to choose between the two tools
- Node types: Input, AI, Output
- The {input} placeholder — how data flows
- Testing and debugging visual flows
- Exporting and documenting workflows
- Map a real business process on paper
- Identify which steps AI can handle
- Set up Flowise or LangFlow account
- Build a sample Feedback Classifier flow
- Build a real 3-step workflow from Week 1 prompt
- Test with 3 real business inputs
- Document with a written process description
- Flowise (free cloud — flowiseai.com)
- LangFlow (free cloud — langflow.org)
- OpenAI or Claude API key (students choose one)
- The difference between a workflow and a tool
- What "vibe coding" means — building without deep expertise
- What an API is and how it works (plain English)
- Getting and safely storing an OpenAI API key
- API cost structure — how to estimate and control spend
- Three deployment paths (Flowise share, HTML form, Replit)
- How to deploy to Netlify in under 2 minutes
- Writing a user guide for non-technical users
- Choose a deployment path based on goals
- Deploy Week 2 workflow as a public tool
- Customise the interface for non-technical users
- Test with a colleague who didn't build it
- Write a 1-paragraph user guide
- Calculate cost per 100 uses
- Flowise public share link
- LangFlow public share link
- OpenAI API (pay-as-you-go, £1–5)
- Netlify or Vercel (free hosting)
- Replit (optional — free tier)
- Trigger → Process → Action systems
- Common business triggers (email, forms, Slack, schedules)
- Zapier vs n8n — which to use and when
- Building a Zap step by step
- Connecting to OpenAI inside Zapier
- Why safeguards matter — and how to add them
- Monitoring automation health and costs
- How to write a Runbook for your team
- Calculating and presenting ROI
- Plan your automation on paper (trigger → action)
- Set up Zapier account and create first Zap
- Connect trigger source to AI step to action
- Test with 5+ real inputs
- Add a human review safeguard
- Monitor for one week — track runs, errors, cost
- Write Runbook for team handover
- Calculate total weekly time saved
- Zapier (free tier — 100 tasks/month)
- n8n (free self-hosted, optional)
- OpenAI API (via Zapier action)
- Google Sheets / Gmail / Slack (as destinations)
Full tool list
How students are assessed
- Week 1: Prompt Card — documented and tested prompt
- Week 2: Workflow export — Flowise or LangFlow, with test evidence
- Week 3: Deployed tool — live URL and user guide
- Week 4: Automation + Runbook — live in Zapier/n8n, with impact metrics
- Prompt produces 4/5 quality output consistently
- Workflow runs on real data without errors
- Tool is accessible by a non-technical user
- Automation runs without manual input for 1 week
- Student can explain every component they built
- ROI is calculated and documented
About the course creator
AI practitioner and course designer specialising in practical AI education for business professionals. This course is built around the principle that the best way to learn AI is to build something real — not watch slides.
You are going to build AI tools. This week.
Not just play with ChatGPT. Actually build — prompts, workflows, deployed apps, and automations running 24/7 in your business. No coding background needed.
One deliverable per week. All of them real.
AI Landscape + Prompt Engineering
Understand what AI can (and can't) do. Learn the structured 5-Part Prompt System that makes AI outputs consistent and professional.
AI Workflows + LangFlow
Stop thinking in single prompts. Build multi-step AI processes visually in LangFlow — no code, just drag, connect, and test.
"Vibe Coding" — Build an AI Tool
Turn your workflow into an app your team can actually use. Deploy it with a link you can share today. No deep coding needed.
Automation + Real Integration
Build trigger-based automations in Zapier or n8n. Connect to your real business tools. Let it run 24/7 while you do other things.
What you need
Total expected cost: £0–30 for the entire 4 weeks. All free tiers are enough to learn and build. Pick either Flowise or LangFlow for Week 2 — you don't need both.
What is AI, really?
Outcome: You understand how AI works — and why that makes you a better prompterBefore you write a single prompt, you need to understand what you're talking to. Not at a PhD level — but enough to stop treating AI like magic and start treating it like a tool with known strengths, known weaknesses, and a very specific history. That understanding is what separates people who get great results from people who don't.
The one thing most people get wrong
Here's the most important thing you will learn today:
It is trained to complete text.
During training, the model sees millions — sometimes billions — of examples of text and learns one thing: given everything I've read so far, what word comes next?
That's it. That's the core mechanism. Everything else — the ability to write emails, answer questions, code software, analyse documents — is a consequence of doing that one thing extremely well, at enormous scale.
Think of the base AI model as a student who has read every book in the library — but has never had a conversation. Technically brilliant. Encyclopaedically well-read. But if you ask it something, it might just continue the sentence rather than actually help you.
That raw model is called the base model. It takes two more stages of training before it becomes the helpful assistant you're used to. More on that shortly.
How AI actually reads your words
Here's something that surprises most people: AI doesn't read text. It reads numbers. Before your words reach the model, they go through a transformation pipeline. Understanding this pipeline explains a lot about why AI behaves the way it does.
Your text gets split into chunks
The model doesn't see words — it sees tokens. A token is roughly a word, part of a word, a punctuation mark, or a space pattern. The model splits your input before it does anything else.
Example — the sentence "The cat sat on the mat." becomes:
Longer or unusual words get split further. "ResearchBuddyAI" might become:
The three stages that turned a text predictor into an assistant
The model that comes out of the tokenisation/embedding/prediction pipeline is not the ChatGPT or Claude you use. It's a raw text completion machine — useful, but not helpful. Three training stages transform it into an assistant.
In pre-training, the model reads an enormous amount of text — books, websites, code, Wikipedia, scientific papers — and learns to predict the next token. It does this billions of times, adjusting its internal weights each time it gets something wrong.
By the end of pre-training it has learned:
Ask the pre-trained base model "Explain AI simply" and it might respond: "Explain AI simply. This topic discusses the fundamentals of..."
It's continuing the text — not helping you. Pre-training = learn language. It does NOT yet = learn to be helpful.
Pre-training is why the model has broad knowledge across almost every topic. But it also means that without further training, it would just complete sentences — not actually answer your question. The next two stages fix that.
After pre-training, the model goes through Supervised Fine-Tuning. Human trainers write thousands of examples of ideal conversations:
The model is trained on thousands of these pairs. It learns: when a user asks a question, I should respond helpfully.
SFT is why giving the model a clear instruction works — it was specifically trained to follow instructions. The clearer and more specific your instruction (your TASK in the 5-Part Framework), the better it performs this trained behaviour.
RLHF — Reinforcement Learning from Human Feedback — is the stage that teaches the model which responses humans actually prefer. SFT taught it to follow instructions. RLHF teaches it to be better at it.
Human raters compare two responses to the same prompt and say which is better:
"Machine learning is an algorithmic paradigm in which statistical models are constructed from training data through gradient-based optimisation..."
"Machine learning means computers learn patterns from data instead of being told rules. Like a child learning to recognise cats — not from a definition, but from seeing thousands of cats."
A reward model learns these human preferences. The main model is then trained to produce responses that score higher — making it progressively better at giving the kind of answers humans prefer.
RLHF is why the model tends to be clear and helpful by default. But it also means the model has preferences — it tends toward certain styles and formats. When your prompt is vague, the model falls back on those trained defaults. When your prompt is specific (ROLE + CONTEXT + CONSTRAINTS), you override those defaults with your own requirements. That's the entire point of the 5-Part Framework.
Why all of this matters for how you prompt
You now understand the full pipeline. Here's how each stage explains something you'll experience every day:
| What you notice | Why it happens | What to do about it |
|---|---|---|
| AI gives vague, generic answers | Vague prompt → model falls back on RLHF defaults | Use ROLE + CONTEXT to override defaults |
| AI states something confidently but it's wrong | Prediction-based, not fact-lookup. Trained on imperfect data | Always verify facts. Ask it to cite or caveat claims |
| AI doesn't know what happened last week | Training has a cut-off date. It learned from past text | Provide current information in your prompt's CONTEXT |
| AI sounds too formal / too casual | Default RLHF style is "professionally helpful" | Specify your exact tone in CONSTRAINTS |
| AI gives a much longer answer than you needed | Default training rewards thoroughness | Set a word limit in CONSTRAINTS |
| AI doesn't know your business, your clients, your context | It only knows what you tell it in each conversation | Put that information in CONTEXT every time |
The three tools you'll use in this course
| Tool | Best for | Cost | Character |
|---|---|---|---|
| ChatGPT | General tasks, code, structured output | Free / £20 per month (Plus) | Direct, fast, widely integrated |
| Claude | Long documents, nuanced writing, analysis | Free / £18 per month (Pro) | Careful, thorough, great at detail |
| Gemini | Google Workspace tasks | Free with Google account | Connected to Google services |
Start with ChatGPT or Claude — both free tiers are enough for Weeks 1 and 2.
Check your understanding
See the training stages in action
Sign up for ChatGPT and Claude (both free). Try these two prompts in both tools and observe the difference:
- Prompt 1 (vague): "Explain machine learning."
Notice: How long is it? How formal? Does it assume your level? - Prompt 2 (specific): "Explain machine learning in 3 sentences for a non-technical marketing manager who just needs to understand the concept well enough to work with AI tools."
Notice: How different is the response? - Now ask each tool: "What was the most significant business news story from yesterday?"
Notice: What do they say? What does this tell you about their training cutoff?
You've just seen RLHF defaults (Prompt 1) vs. overriding them with specifics (Prompt 2), and the pre-training cutoff limitation (Prompt 3). These three observations will change how you prompt from now on.
The 5-Part Prompt Framework
Outcome: You write prompts that work the first time — every timeMost people type questions into AI like they're texting a friend. That's why they get mediocre results. The difference between a £5-an-hour assistant and a £500-an-hour consultant is how you brief them. This lesson teaches you to brief AI like a pro — and then challenges you to spot bad prompts in the wild.
Why most prompts fail
Here's what most people type: "Write me a marketing email."
Here's what AI actually needs to write something useful:
- Who is writing it? (Your company, your expertise, your tone)
- Who is receiving it? (What do they care about? What stage of the funnel?)
- What is the goal? (Click a link? Book a call? Buy something?)
- What constraints apply? (Length, formality, number of CTAs)
- What should the output look like? (Subject + body? Just body? HTML?)
Without those answers, AI guesses. Guesses are generic. Generic wastes your time. The 5-Part Framework eliminates guessing.
The 5-Part Framework — every part earns its place
Tell AI who it should be. This sets expertise, tone, and how it frames its answers. The more specific, the better.
"You are a senior copywriter who specialises in B2B SaaS email campaigns with 10+ years of experience."Give background. Don't make AI guess your situation. The more specific detail you provide, the more relevant the output.
"We're a 10-person agency launching a project tracker for marketing teams. Customers are marketing managers at 50–200 person companies who currently use spreadsheets."Be precise about what you want done. Strong verbs: write, analyse, summarise, rewrite, compare, extract, generate, classify.
"Write a 3-email onboarding sequence welcoming new free trial users and guiding them to activate their first project within 48 hours."This is where most people skip — and regret it. Tell AI exactly what to avoid, what limits apply, what tone to use.
"Keep each email under 150 words. No jargon. One clear CTA per email. Warm, helpful tone — not corporate, not pushy."Say exactly how the result should appear. A list? A table? A document? If you don't specify, you'll get whatever AI finds easiest — rarely what you need.
"Format: Subject line / Preview text (40 chars max) / Email body. Label each Email 1, 2, 3."Real examples — see the difference yourself
Each example below shows the same goal tackled first with a weak prompt, then with the 5-Part Framework. Read both carefully. You'll start seeing exactly what's missing.
"Dear Customer, We sincerely apologise for the delay in your order. We are working hard to resolve this issue and will update you shortly. Thank you for your patience. Regards, The Team."
- No role — reads like a robot wrote it
- No context — doesn't know how late, what product, who the customer is
- No constraints — could be any length, any tone
- No output format — just a blob of text
Context: A customer ordered a birthday gift. It's 3 days late. They emailed us once already — frustrated but polite. The gift was £80 worth of chocolates.
Task: Write a personal apology email that retains their loyalty and makes the situation right.
Constraints: Under 120 words. Warm and human, not corporate. Offer either a discount code OR expedited shipping — not both. Use the customer's name placeholder [Name].
Output: Subject line + email body only.
"Hi [Name], I'm so sorry about this — a birthday gift arriving late is genuinely awful, and you deserved better from us. I've personally flagged your order and [expedited shipping / a 20% discount code: SORRY20] is on its way to you now. Thank you for handling this so graciously. — Sarah, Customer Success"
- Feels human — specific role informed the tone
- Addresses the real situation — the birthday, the amount
- Stays under word limit from the constraints
- Exactly the format you asked for
A list of 5 utterly generic ideas: "Share a personal story," "Post a tip about your industry," "Celebrate a team win," "Ask your audience a question," "Share a lesson you learned." — You could have typed those yourself in 10 seconds.
- No role — who's writing these? A CEO? A freelancer?
- No context — what industry, what audience, what goal?
- No constraints — how many? What style? What to avoid?
- No output format — just ideas or full drafts?
Context: I'm a UX consultant with 8 years of experience. My audience is product managers and startup founders. I post twice a week. My goal is to generate inbound enquiries — not just likes.
Task: Generate 5 LinkedIn post concepts for this week focused on common UX mistakes that cost companies money.
Constraints: Each concept should feel original — no "5 tips" listicles. At least one should use a real-world failure as the hook. No motivational fluff. Max 30 words per concept description.
Output: Numbered list. Each entry: Hook sentence + 1-line description of the angle.
5 sharp, specific post concepts — with hooks like "We redesigned a checkout flow and killed £40k in revenue. Here's what we missed." Each one usable, specific to your audience, and with a clear angle.
- Role told AI the content style and platform expertise
- Context gave audience, goal, and posting frequency
- Constraints killed the generic listicle format
- Output format made the response immediately actionable
A vague response: "Based on the information provided, this lead appears to have potential. They work in technology and seem interested in your services. I would recommend following up to learn more about their needs." — Useless. You already knew that.
- No role — AI doesn't know your sales criteria
- No context — what counts as a "good" lead for you?
- No task specifics — what does "analyse" mean here?
- No output format — you need a score, not an essay
Context: We sell project management software at £299/month. Ideal customers are marketing agencies with 10–50 staff who currently use spreadsheets. We close deals that have: budget sign-off, a current pain point, and a decision-maker as the contact.
Task: Score this incoming lead from 1–10 and explain your reasoning. Flag any missing information we should ask for on the follow-up call.
Constraints: Be direct. Don't recommend following up with every lead — say "not worth pursuing" if the score is below 4. Max 150 words total.
Output: Score: X/10 | Reason: [2 sentences] | Missing info: [bullet list] | Recommended next step: [one sentence]
Score: 7/10 | Reason: Contact is a Head of Marketing at a 22-person agency — fits ICP. Mentioned "drowning in spreadsheets" which is a strong pain signal. | Missing info: Budget authority, timeline, current tool spend | Next step: Book discovery call, lead with the spreadsheet comment.
- Context defined exactly what a good lead looks like
- Task asked for a specific score, not a vague assessment
- Constraints forced direct, actionable language
- Output format means you can scan 50 leads in minutes
A paragraph that retells everything you just said — essentially your notes reorganised into different sentences. Nothing was extracted. Nothing was prioritised. You still have to read the whole thing.
- No role — who is this summary for? What do they care about?
- No task specifics — summarise into what? A paragraph? Action items?
- No constraints — how long? What to include or skip?
- No output format — just prose, not usable structure
Context: This is a 45-minute client kick-off meeting for a website redesign project. The output goes to three people: the client (non-technical), the project manager, and the lead developer.
Task: Extract all decisions made, all action items with owners and deadlines, and any risks or open questions raised.
Constraints: Summary should be under 120 words. If a deadline wasn't mentioned, write "deadline TBC — needs confirmation." Don't include small talk or filler. Flag any conflicting instructions as ⚠️.
Output: Use this exact structure:
SUMMARY (3 bullets max) | DECISIONS | ACTION ITEMS (Task · Owner · Deadline) | OPEN QUESTIONS | RISKS
A crisp, scannable summary your team can act on immediately — with named owners on every action item, flagged conflicts, and nothing that wastes anyone's time.
- Role told AI exactly what to prioritise and how to think
- Context gave audience — critical for what to include or omit
- Constraints handled the "deadline TBC" edge case automatically
- Output format is reusable across every future meeting
Build your own prompt live
5-Part Prompt Builder
Fill in each section below. Your prompt assembles at the bottom in real time — copy it straight into ChatGPT or Claude.
Now it's your turn — spot the mistakes
Below are 5 real prompts that professionals actually use. Each one has problems. Your job: identify what's missing or broken in each prompt, then submit your answers. You'll see how your thinking compares across all 5 areas.
Select all the problems you can spot in each prompt. There may be more than one correct answer per question.
The refinement loop
Your first prompt won't be perfect. That's expected — and fine. Here's how to improve it systematically:
Run it with real input
Don't test with a made-up example. Use something from actual work this week.
Rate the output 1–5
Ask yourself honestly: would I use this as-is, or does it need major rework?
Diagnose which part failed
Too generic? → ROLE. Wrong topic? → CONTEXT. Wrong format? → OUTPUT. Too long/wrong tone? → CONSTRAINTS.
Change one thing at a time
Fix only the section that's wrong. This way you know exactly what made the difference.
Save your best version
When you hit a 4 or 5, document it as a Prompt Card. That's your template for every future use of that task.
Build your AI Work Assistant
Outcome: A prompt you'll actually use in your job next weekThis is your Week 1 project. You're going to pick one repetitive task from your real work and build a 5-Part prompt that handles it consistently and well. Then you're going to test it three times with real inputs. By the end, you'll have something you can use on Monday.
Step 1 — Choose your task
Pick something you do at least once a week. The best candidates are tasks that are:
- Repetitive (same structure every time, different content)
- Writing-heavy (drafts, summaries, responses, reports)
- Time-consuming but not deeply strategic (you shouldn't need to think hard about format)
- Writing follow-up emails after meetings or proposals
- Summarising customer feedback or support tickets
- Drafting social media posts from a brief
- Creating first drafts of proposals or SOWs
- Writing weekly update messages to clients or teams
- Generating meeting agendas from a topic list
- Responding to common customer enquiries
- Analysing competitor content or positioning
Step 2 — Write your prompt
Use what you learned in Lesson 2. Here's a template you can copy and fill in:
Step 3 — Test three times
Don't just test with a perfect, easy example. Use real inputs from your work:
Easy test
A simple, typical version of the task. This confirms the prompt works at all.
Real test
An actual example from last week's work. Does it produce something you'd use?
Edge case test
An unusual version — a difficult customer, a complex situation, an edge case. Does it break?
For each test, note: what score (1–5) would you give the output? What would you change?
Step 4 — Document it as a "Prompt Card"
A Prompt Card turns your prompt into a reusable tool. Here's the format:
Week 1 completion checklist
- ✓Signed up for ChatGPT or Claude (free tier)
- ✓Identified my repeating task
- ✓Written my 5-Part prompt
- ✓Tested it 3 times with real inputs
- ✓Refined based on test results
- ✓Saved as a Prompt Card
- ✓Used it in real work at least once
Your Prompt Card
Submit (or share in your cohort channel) a completed Prompt Card including:
- Your full 5-Part prompt
- Three test outputs with your quality scores
- What you refined and why
- Your time-saved estimate per week
This prompt card is the foundation of Week 2. You'll use it as your starting point when building your first LangFlow workflow. Don't skip saving it.
AI Workflows + Visual Building
- 🔒 Thinking in Systems
- Flowise & LangFlow: Your First Flow
- Build & Test Your Real Workflow
Flowise & LangFlow
- 🔒 Thinking in Systems
- Flowise & LangFlow: Your First Flow
- Build & Test Your Real Workflow
Build & Test Your Workflow
- 🔒 Thinking in Systems
- Flowise & LangFlow: Your First Flow
- Build & Test Your Real Workflow
"Vibe Coding" — Build an AI Tool
- 🔒 From Workflow to App
- APIs Without the Jargon
- Deploy Your Tool
APIs Without the Jargon
- 🔒 From Workflow to App
- APIs Without the Jargon
- Deploy Your Tool
Deploy Your Tool
- 🔒 From Workflow to App
- APIs Without the Jargon
- Deploy Your Tool
Trigger → Action Systems
- 🔒 Trigger → Action Systems
- Build Your Automation in Zapier
- Monitor, Scale & Ship
Build in Zapier
- 🔒 Trigger → Action Systems
- Build Your Automation in Zapier
- Monitor, Scale & Ship
Monitor, Scale & Ship
- 🔒 Trigger → Action Systems
- Build Your Automation in Zapier
- Monitor, Scale & Ship
Tools & Resources
- 🔒 Full tool list, links & setup guides
- Cost comparison by week
- Honest gaps & what to learn next
Glossary
- 🔒 Plain-English definitions for every term
- From API to {input} placeholder
- 30+ entries