# AI Native Course - full course text > AI that works for you - From Automation to Leverage. Free, hands-on AI training for product, growth, design and engineering teams. Last reviewed: August 2026. Source: https://ai-course.dev/ · Licence: free to read and cite with attribution to ai-course.dev. ## Module 1: Understanding LLMs & Choosing Your AI Stack URL: https://ai-course.dev/module/understanding-llms Duration: ~35 min · Category: Orientation Before you build anything with AI, you need to understand what you're actually working with. This module covers how large language models work (in plain English), how the major models compare, and the key concepts - projects, skills, MCPs, direct connections - that will shape every decision you make going forward. ### Lesson 1. How LLMs actually work (plain English) A large language model is not a search engine, a database, or a calculator. It's a pattern-completion system trained on an enormous amount of text. Given an input, it predicts the most statistically likely continuation - shaped by billions of parameters learned during training. What this means practically for a PM: - It has no memory between conversations unless you give it one - It doesn't "know" things the way a database does - it has learned patterns and associations - It can be confidently wrong (hallucination) because it optimises for plausible, not true - The quality of its output is heavily shaped by the quality and structure of its input - It reasons better when given time to think (chain-of-thought) vs asked for instant answers - It has a context window - a limit on how much text it can hold in one conversation Understanding these properties is what separates someone who gets frustrated with AI from someone who knows how to get consistent, high-quality output. 2. LLM Comparison: Claude vs ChatGPT vs Gemini Dimension | Claude (Anthropic) | ChatGPT (OpenAI) | Gemini (Google) | Best known for | Long-form reasoning, nuanced writing, careful instruction-following | Breadth of capability, largest ecosystem, most integrations | Google Workspace integration, multimodal, search-connected | Strengths for PMs | Deep analysis, research synthesis, long documents, structured outputs | Versatility, code gen, image understanding, GPT Store, brainstorming | Real-time web access, native Google Docs/Sheets, data-heavy tasks | Weaknesses | No native real-time web in base mode, smaller plugin ecosystem | Can be verbose, over-hedges, cost at scale | Less nuanced on complex reasoning, enterprise privacy concerns | Context window | Very large (hundreds of thousands of tokens) - whole documents and codebases | Large, and larger still on the reasoning tiers | The largest of the three - up to ~1M tokens on the Pro tier | Best PM use case | Research synthesis, long briefs, structured analysis, automation prompts | Brainstorming, code, quick tasks, image analysis | Google Workspace workflows, real-time research, Sheets analysis | Pricing | Claude Pro $20/mo or API | ChatGPT Plus $20/mo or API | Gemini Advanced $20/mo or API | Recommendation: For most growth PM workflows - research synthesis, experiment briefs, automation prompts, structured analysis - Claude is the strongest default. For brainstorming and quick tasks, ChatGPT is excellent. For anything living inside Google Workspace, Gemini wins on integration. Most power users have 2–3 and route by strength. Think in tiers, not version numbers Model names change every few months; the shape of the choice doesn't. Every major provider ships roughly three tiers, and picking the tier is 90% of the decision: Tier | What it's for | Rule of thumb | Frontier / reasoning | Hard analysis, long documents, multi-step planning, agentic work | Slowest and priciest - use when being wrong is expensive | Balanced / default | Everyday drafting, synthesis, summarising, most PM work | Your default in the chat UI | Fast / cheap | Tagging, extraction, classification, high-volume automation | Use inside loops and pipelines where you run it 1,000 times | When you read a benchmark or a launch post, translate it into a tier before you change anything in your workflow. If your prompt only works on one specific version, the prompt is fragile - not the model. 3. Projects vs Skills: What's the difference? A Project (Claude Projects, ChatGPT Custom GPTs) is a persistent context container. It lets you give the AI a standing system prompt ("You are always working as a growth PM at your company…"), upload reference docs it can always access, and maintain a consistent role and tone across sessions. Think of a Project as a pre-briefed AI colleague dedicated to a specific area of your work. Examples for a Growth PM: "Growth Workspace" (role, ICP, OKRs, experiment log) · "User Research" (framework, taxonomy) · "Content & Copy" (brand voice, tone examples). A Skill means different things in different tools - in agent frameworks it's a reusable instruction set, in Microsoft Copilot it's a plugin capability, in general usage it's a saved prompt template. The mental model: Projects are about persistent context. Skills are about reusable actions. | Project | Skill | What it is | Persistent context container | Reusable action / prompt template | Sets up | Who AI is, what it knows, how it should behave | What AI should do when triggered | Analogy | Onboarding a colleague with full context | Giving that colleague a standard operating procedure | PM use case | "Always know I'm a growth PM at my company" | "When I say 'write a debrief', follow this structure" | 4. Why MCP instead of a direct connection? Direct API: you call an external service's API directly in your code or n8n flow. Works for simple linear workflows, full control over the call - but every new AI tool needs custom integration code, and there's no standardised way for the AI to discover available tools. MCP (Model Context Protocol): a standardised protocol (open-sourced by Anthropic, widely adopted) that sits between the AI and your tools. Write the integration once as an MCP server; any MCP-compatible AI client (Claude, Cursor, Windsurf, etc.) can use it. The AI dynamically discovers tools, tools self-describe, and switching models doesn't break integrations. The key insight: Direct API = you decide when to call the tool. MCP = the AI decides when to call the tool. MCP shifts integration from "developer wiring things together manually" to "AI dynamically using tools as needed." That's the difference between an AI that answers questions and an AI that takes actions in your systems. Situation | Use | One-off automation, linear flow, existing n8n node | Direct connection / n8n | Want any AI to always be able to query your internal data | MCP | Building something other tools or teammates will use | MCP (standardised) | Prototyping quickly, complexity not justified | Direct API | Want AI to autonomously decide when to pull data | MCP | ### Prompts #### Which AI tool should I use for this? I need to complete the following task: [DESCRIBE TASK IN DETAIL]. My context: I am a Growth PM. My most-used tools are Claude, ChatGPT, and Gemini. Recommend which AI tool is best suited for this specific task and why. Consider: reasoning quality, context window needs, integration requirements, real-time data needs, and output format. Be specific about which tier to use (frontier reasoning tier vs fast/cheap tier) rather than a version number, since versions change every few months. #### Set up my AI Project Help me set up a Claude Project (or ChatGPT Custom GPT) for the following use case: [DESCRIBE USE CASE - e.g. "growth PM work at your company"]. Write the system prompt I should paste into the project. It should include: 1. My role and context 2. The tone and format I want responses in 3. What I'll primarily use this project for 4. Standing instructions that should always apply 5. What the AI should always avoid Keep the system prompt under 500 words and make it immediately usable. #### Should I build an MCP or use a direct connection? I want to connect AI to the following tool/system: [DESCRIBE TOOL - e.g. "our experiment tracker in Notion", "our Contentful CMS", "our internal analytics API"]. My use case is: [DESCRIBE USE CASE]. Help me decide: should I build an MCP server, use a direct API connection, or use an existing integration (e.g. n8n node)? Consider: frequency of use, whether multiple AI tools need access, complexity of the integration, and whether the AI needs to autonomously decide when to query it. Give a clear recommendation with reasoning. ### Exercise 🛠 Exercise 1 - Run the same prompt across 3 models. Pick a real task from your week (e.g. "write a hypothesis for why our activation rate dropped"). Run the exact same prompt in Claude, ChatGPT, and Gemini. Compare output quality, structure, tone, and what each missed. Note which one you'd use for this task type going forward. 🎯 Exercise 2 - Set up your first AI Project. Create a Claude Project (or ChatGPT Custom GPT) for your growth PM work. Use the "Set up my AI Project" prompt to generate the system prompt. Test it with 3 real tasks. Notice the difference vs a blank conversation. 💡 Exercise 3 - Map your tools to the right model. List your 5 most common AI use cases. For each one, decide: Claude, ChatGPT, or Gemini - and why. This is your personal routing guide. Revisit it in 4 weeks and see if your answers have changed. ## Module 2: Prompt Engineering That Actually Works URL: https://ai-course.dev/module/prompt-engineering Duration: ~35 min · Category: Foundation Most PMs use AI like a search engine. This module rewires that. You'll learn systematic prompting techniques - chain-of-thought, few-shot, role framing, output structuring - with examples grounded in real growth work. ### Lesson The 5 layers of a great prompt - Role - who the model should be (“You are a senior growth PM…”) - Context - the situation, constraints, audience - Task - the single, specific outcome you want - Constraints - format, length, tone, things to avoid - Examples - 1–3 worked examples (few-shot) Chain-of-thought Adding think step by step forces the model to surface its reasoning before concluding. This catches logical errors and dramatically improves analytical tasks. Output formatting Ask explicitly for JSON or a markdown table when the output will be parsed or pasted into Notion. Structure is leverage - it makes downstream automation trivial. Few-shot prompting One great example beats a paragraph of instructions. For recurring PM tasks, keep a small library of input → output pairs you can paste into the prompt. Part 2 · Advanced Techniques Part 2: Advanced Techniques System prompts - your persistent instruction layer A system prompt is a set of instructions that runs before every message in a conversation. It's the difference between starting every chat from scratch and starting every chat with a fully briefed AI that already knows your role, your standards, and your context. In Claude Projects and ChatGPT Custom GPTs, the system prompt is set once and applies to the whole project. In API calls, it's the system parameter. Either way, it's the highest-leverage prompt you'll ever write - because it compounds across every subsequent interaction. What belongs in a system prompt: - Your role and company context - The tone and format you always want - What the AI should never do - Standing rules that apply to every task ("always return JSON", "always include a confidence score", "never recommend solutions that require engineering if a no-code option exists") A strong system prompt turns a general-purpose model into a specialist that knows your world. Multi-turn conversation design Most people use AI as a single-shot vending machine: one prompt in, one response out, start over. The real leverage is in conversation design - structuring a multi-turn dialogue where each turn builds on the last. Patterns that work: - Diverge then converge: Turn 1 - generate 10 options. Turn 2 - filter to the 3 best. Turn 3 - develop the winner in full. - Critique loop: Turn 1 - generate output. Turn 2 - "critique this output against [criteria]". Turn 3 - "rewrite based on that critique." - Interrogation first: Turn 1 - "ask me the questions you need answered before writing this brief." Turn 2 - answer them. Turn 3 - write the brief. Output quality is dramatically higher when AI has asked its own clarifying questions. - Role reversal: Turn 1 - "you are a sceptical CPO reviewing this proposal. What are your top 3 objections?" Turn 2 - address the objections. Turn 3 - rewrite the proposal. Prompt chaining for complex workflows A prompt chain is a sequence of AI calls where the output of each step becomes the input of the next. This is the architecture behind most serious AI workflows - including your insight report automation. Example chain for a growth experiment brief: - "Given this metric drop: [DATA], generate 8 hypotheses" → output: hypothesis list - "From this list, identify the 2 highest-leverage hypotheses based on: [CRITERIA]" → output: 2 hypotheses - "Write a full experiment brief for hypothesis #1 using this template: [TEMPLATE]" → output: brief Each step is simple. The chain is powerful. This is also how you break tasks that are too complex for one prompt into reliable, auditable steps. Context engineering - the skill after prompting Prompting is what you type. Context engineering is everything else the model can see when it answers: project files, retrieved documents, memory, tool outputs, conversation history. As models got better at following instructions, the bottleneck moved from wording to context. - Curate, don't dump. A long context full of irrelevant material measurably degrades answers. Ten pages of the right thing beats 200 pages of everything. - Retrieval over paste. If the source changes (docs, tickets, analytics), connect it - a project file, a connector, or an MCP server - instead of pasting a stale copy. - Memory is a design decision. Decide what should persist across sessions (role, ICP, tone, definitions) and what should be per-task. Persist the stable things in a Project system prompt. - Watch for context rot. In long threads, early instructions get diluted. Restate the goal, or start a fresh thread with a short handover summary. - Structure the context. Label sections (, , ) so the model can reference them precisely. A practical test: if you handed the same context pack to a smart new colleague, could they do the task? If not, no prompt wording will save you. Temperature and model selection intuition You don't need to understand transformer architecture. You do need intuition about two things: Temperature (in API settings) controls randomness. Low temperature (0–0.3) = consistent, conservative, deterministic - best for structured outputs, JSON, classification. High temperature (0.7–1.0) = more creative, more varied - best for brainstorming, copy variants, ideation. Most chat interfaces don't expose this directly, but some tools do. Model selection: Bigger ≠ always better. For simple classification or extraction tasks, the fast/cheap tier of any provider is far quicker and cheaper than the frontier tier. Use the most powerful model for: complex reasoning, long document analysis, nuanced judgment calls. Use lighter models for: tagging, classification, simple transformations at scale. The PM intuition: match model power to task complexity. Over-specifying wastes cost and latency. Under-specifying wastes output quality. ### Prompts #### Growth brief writer You are a senior Growth PM at a PLG SaaS company. Context: We're running weekly experiments on activation. I need a brief for a new test. Task: Write an experiment brief. Constraints: ≤ 250 words. Use sections: Hypothesis, Metric, Variant, Success criteria, Risks. Example structure: - Hypothesis: If X then Y because Z - Metric: Day-7 activation rate (primary) Now write the brief for: [paste idea here] #### Hypothesis generator You are a growth strategist. A key metric dropped: [describe metric, magnitude, timeframe]. Generate 6 hypotheses for why this happened. Return as a markdown table with columns: | # | Hypothesis | Funnel stage | How to validate | Confidence (L/M/H) | Think step by step before answering. #### Prompt improver Rewrite the following weak prompt using the 5-layer framework (Role, Context, Task, Constraints, Examples). Explain in 2 lines what changed and why. Weak prompt: """ [paste prompt] """ #### Critique loop I've just written the following [DOCUMENT TYPE - brief/spec/analysis]: [PASTE DOCUMENT] Step 1: Critique this document against these criteria: [LIST CRITERIA - e.g. clarity, completeness, actionability, evidence quality]. Be direct and specific. Don't soften the feedback. Step 2: Rewrite the document addressing every critique. Keep what works, fix what doesn't. Return the rewritten version in full. #### Interrogation-first brief I need you to write a [DOCUMENT TYPE] about [TOPIC]. Before you write anything, ask me the 5–7 questions you need answered to produce the best possible output. Ask them one at a time and wait for my answer before asking the next. Only start writing after you have all the answers. ### Exercise Pick a real recurring PM task (weekly update, experiment brief, NPS digest). Build a reusable prompt using the 5-layer framework. Add think step by step. Run it twice - once with your old prompt, once with the new one. Note the quality difference and what specifically improved. ## Module 3: AI for User Research & Insights URL: https://ai-course.dev/module/ai-for-user-research Duration: ~40 min · Category: Research You already ship insight reports at your company. This module is about 10x-ing that process - synthesizing interviews at scale, auto-tagging feedback, building living insight repositories, and extracting signal from noise. ### Lesson The real bottleneck It's not collection. It's synthesis. AI is uniquely good at compressing 50 transcripts into a JTBD map in minutes. Auto-tagging feedback Define a taxonomy once. Have AI classify every new piece of feedback into it with a confidence score. Suddenly qualitative data becomes a quantifiable signal you can chart. Living insight repository Pipe tagged feedback → weekly digest → Notion database. Insights compound. Patterns surface that no single human would notice. Reuse the architecture you already have Your Contentful translation agent is the same shape: input → AI → structured output → destination. Insight automation is a sibling, not a new beast. Building a living insights repository The real unlock from AI-powered research isn't individual syntheses - it's a compounding knowledge base. Every month: AI processes new feedback → structured summaries flow into a Notion database → any PM can query "what do users say about onboarding friction?" and get an instant synthesis of 6 months of data. The architecture: a Notion database with properties for Theme, Source, Date, Sentiment, Representative Quote. AI populates it. You query it. New team members can get up to speed on user context in an hour instead of a month. The research bottleneck: synthesis, not collection Most teams collect enough data. The bottleneck is turning it into something actionable. AI doesn't replace your judgment about what matters - it removes the labour of organising what exists so you can apply that judgment faster. The PM who synthesises research 10x faster doesn't just save time. They run more research cycles, which means better product decisions, compounding over time. ### Prompts #### Interview synthesizer You are a senior user researcher. Below are N user interview transcripts (separated by ---). Extract: 1. Top 5 Jobs-To-Be-Done (with frequency) 2. Top 5 pain points (with frequency and severity L/M/H) 3. 3 surprising or counter-intuitive findings 4. 5 representative verbatim quotes Return as markdown. Think step by step. Transcripts: [paste here] #### Feedback tagger Classify each feedback item below into our taxonomy. Taxonomy: [Onboarding, Templates, Collaboration, Pricing, Performance, Mobile, Other] For each item return JSON: { "text": "...", "tag": "...", "confidence": 0.0-1.0, "sentiment": "pos|neg|neu" } Return a single JSON array. Items: [paste] #### NPS insight extractor You are a PM analyst. Below are NPS verbatims (one per line). Return: - Overall sentiment breakdown (% pos/neu/neg) - Top 5 themes with count and a representative quote - 3 recommended product actions, ranked by leverage Format: markdown. Be specific, no fluff. Verbatims: [paste] ### Exercise Take real NPS or Intercom data from your company. Run the feedback tagger. Build a theme → count → quote table. Then sketch a weekly automation flow: source → AI → Notion digest. What does the trigger look like? Who reads it on Monday morning? ## Module 4: AI-Powered Growth Workflows URL: https://ai-course.dev/module/ai-powered-growth-workflows Duration: ~35 min · Category: Growth If you're running experiments and tracking growth metrics, this module accelerates every stage of the growth loop using AI - from hypothesis generation to copy testing to experiment debriefs. ### Lesson AI across the growth loop Attract · Activate · Retain · Refer · Revenue. Every stage has high-volume repetitive work AI eats for breakfast: ideation, copy, analysis, debrief. Hypotheses at scale From one signal - a metric drop, a support theme - generate 8 hypotheses in 30 seconds. Then filter ruthlessly. Quantity → diversity → judgment. Copy testing at machine speed Generate 10 variants across emotional angles (curiosity, social proof, urgency, benefit, pain, question). Pick the 2 best. Ship. Measure. Debrief automation A templated debrief after every experiment turns learnings into a searchable knowledge base. Compounding > one-off wins. Section · AI for Data & Analytics AI for Data & Analytics Data is where growth PMs spend a disproportionate amount of time - pulling numbers, writing queries, interpreting charts, explaining results to stakeholders. AI compresses every stage of this loop. Querying data in plain English Tools like Rows AI, PostHog AI, and Claude's data analysis mode let you ask questions in plain English instead of writing SQL or building Looker dashboards. "What's our week-over-week activation rate for users who signed up via the referral channel?" is now a prompt, not a 45-minute analytics task. For existing SQL or analytics work, you can paste a query and ask Claude to: explain it, optimise it, modify it, or translate it to a different syntax. You don't need to become a SQL expert - you need enough fluency to verify AI's output. Interpreting charts and dashboards Screenshot any chart and paste it into Claude with: "You are a growth analyst. What story is this chart telling? What's the most important signal? What would you investigate next?" The output is often a faster and more structured insight than writing it yourself - especially for charts you're seeing for the first time. This is the fastest path from "data dump" to "stakeholder-ready narrative." Anomaly analysis When a metric moves unexpectedly, the bottleneck is usually generating plausible explanations quickly. AI is excellent at this: "Our 7-day activation rate dropped 11% between [DATE A] and [DATE B]. Here are the metrics that moved with it: [DATA]. Generate 6 hypotheses ordered by likelihood. For each: the hypothesis, what supporting signal would confirm it, and what data I'd need to rule it out." This turns a 2-hour investigation kickoff into a 10-minute structured brief. Generating data narratives for stakeholders Paste your weekly/monthly metrics into Claude and ask it to write the narrative: "Here are our growth metrics for this week: [DATA]. Write a 200-word summary for our leadership team. Lead with the most important movement, explain the likely cause, and end with the one metric to watch next week." This is not about removing your judgment - it's about removing the blank-page problem. You edit the narrative, you don't write it from scratch. The critical limit: AI cannot access your data AI works with data you give it. It cannot connect to GA4, Mixpanel, or Looker directly - unless you build an MCP or integration that provides that access (covered in Module 5). Until then, the workflow is: export/copy data → paste into AI → get analysis. It's still dramatically faster than manual analysis, but it's not real-time. The unlock is building integrations (n8n → AI, or an analytics MCP) so data flows to AI automatically instead of manually. That's the automation layer. ### Prompts #### Hypothesis generator You are a growth PM. Signal: [paste data + context]. Generate 8 hypotheses. Return as a markdown table: | # | Hypothesis | Funnel stage | Experiment | Primary metric | Confidence (L/M/H) | Think step by step. #### Copy variant generator Generate 6 email subject lines for: [campaign description, audience]. One per emotional angle: curiosity, social proof, urgency, benefit, pain, question. Each ≤ 55 chars. Return as a markdown table with columns: # | Angle | Subject line | Why it works. #### Experiment debrief You are a growth PM writing a post-experiment debrief. Inputs: [hypothesis, variant, result, p-value, segment cuts]. Return sections: 1. What we shipped (1 line) 2. Result (with confidence) 3. What we learned (3 bullets) 4. Why it (didn't) work (causal reasoning) 5. What to test next (2 specific ideas) #### Metric anomaly investigator You are a senior growth analyst. Our [METRIC] moved by [DIRECTION + AMOUNT] between [DATE A] and [DATE B]. Context: - Product: [PRODUCT] - User segment affected: [SEGMENT or "all users"] - Other metrics that moved at the same time: [LIST] - Recent changes (releases, campaigns, seasonality): [LIST or "none known"] Generate 6 hypotheses for what caused this movement. For each hypothesis: 1. The hypothesis (specific and falsifiable) 2. The signal that would confirm it 3. The data source I'd need to check 4. Confidence level: high / medium / low Order by likelihood. Return as a markdown table. #### Stakeholder metrics narrative You are a growth PM presenting to leadership. Here are this [week/month]'s key metrics: [PASTE METRICS - numbers, % changes, context] Write a 200-word narrative update. Structure: 1. Lead with the single most important movement (positive or negative) 2. Explain the most likely cause in one sentence 3. Call out one risk or thing to watch 4. End with the one metric that matters most next week/month Tone: direct, data-grounded, no fluff. No bullet points - write in prose. ### Exercise Pick a real metric from your product. Generate 8 hypotheses. Rank by impact × ease × confidence. Pick the top 2 and draft briefs. Then generate 6 copy variants for a real campaign. Pick 2 to test. Note which angle you'd never have written alone. ## Module 5: Building AI Agents & Automations URL: https://ai-course.dev/module/building-ai-agents Duration: ~45 min · Category: Engineering You've built an insight report automation and a Contentful translation agent. This module formalizes the mental model - how to think about agents, when to use n8n vs code vs APIs, and how to design reliable multi-step workflows. ### Lesson The 4-step agent pattern - Receive a goal - Use tools (APIs, files, search) - Reason about what's next - Repeat until done or stuck Decision framework n8n - linear logic, existing integrations, non-engineers can edit. Direct API - custom logic, performance-critical, tight integration with your code. Agent framework (LangChain, CrewAI) - multi-step reasoning, dynamic tool selection. Reliability Most AI automations work 80% of the time and fail silently. Add: input validation, output schema checks, retry with backoff, human-in-the-loop for high-stakes paths, observability. Prompts as code Version them. Test edge cases. Document the expected output contract. A regression suite for prompts is not overkill - it's table stakes once a flow runs in prod. Prompt templates as code (in practice) In automations, treat your prompts like code. Version them. Test edge cases. Document the expected input format and output schema. A prompt that works 90% of the time is a function with a 10% bug rate - which is fine in one-off use and unacceptable in automation. Specific practices: - Save your automation prompts in a version-controlled doc (Notion, GitHub, anywhere with history) - Write test cases: what happens when the input is empty? When it's in a language you didn't expect? When it's 10x longer than normal? - Specify the output format as a schema, not just a description: "Return a JSON array of objects with keys: theme (string), count (integer), confidence (high|medium|low)" - When a prompt starts producing worse output over time, treat it like a bug: investigate, iterate, document the fix ### Prompts #### Agent design prompt Spec an automation for: [recurring task]. Return: - Trigger (schedule / webhook / manual) - Inputs (sources, format) - AI step (model, prompt sketch) - Output format (JSON schema if possible) - Destination (where output lands) - Error handling (what could fail, fallback) - Human-in-the-loop touchpoints #### n8n workflow designer Design an n8n workflow for: [task]. List nodes in order. For each: node type | function | key config | error path. Highlight any node where I should add a "If error → notify Slack" branch. #### Automation audit Here is an automation I built: [paste flow description]. Audit it. Find: 1. Single points of failure 2. Missing error handling 3. Prompt reliability issues (vague outputs, no schema) 4. Scalability concerns (rate limits, cost at 10x volume) Return prioritized fixes. ### Exercise Spec a real recurring task as an automation using the agent design prompt. Then audit your existing insight report automation for edge cases and ship a fix for the highest-risk one. ## Module 6: MCPs: Build Your Own URL: https://ai-course.dev/module/mcps-build-your-own Duration: ~45 min · Category: Advanced Many teams are now building custom MCPs for their internal tools. Now it's your turn - understand them deeply, learn the anatomy, and build one from scratch. MCPs are the missing layer between AI and your internal tools. ### Lesson What MCP is Model Context Protocol - an open standard for connecting AI assistants to your tools, data, and prompts. Think USB-C for AI. Anatomy of an MCP server - Tools - actions the model can invoke (query DB, call API) - Resources - data the model can read (files, records) - Prompts - reusable templated workflows - Transport - stdio (local) or HTTP/SSE (remote) MCP vs n8n n8n is scheduled workflow. MCP is persistent capability. With MCP, your AI can always reach your systems on demand - no trigger needed. ~50 lines to a working MCP The TypeScript SDK lets you spin a server with one tool in under 50 lines. Start there. Add tools as you discover what you keep copy-pasting into Claude. The MCP ecosystem - what already exists Before building a custom MCP, check what's already available. A growing ecosystem of pre-built MCP servers covers: Notion, Linear, GitHub, Slack, Google Drive, HubSpot, Contentful, and many more. If your tool is already covered, you can connect it to Claude in minutes - no building required. The value of building your own MCP is for internal tools and proprietary data that don't have a pre-built integration: your experiment tracker, your internal analytics API, your custom data warehouse. That's where the leverage is. Check the official MCP registry at modelcontextprotocol.io/registry before building from scratch. ### Prompts #### MCP tool designer Design 5 tools for a company-internal MCP server. Company: [describe]. Team: [Growth PM]. For each tool return: - name (snake_case) - description (1 line, model-facing) - inputs (JSON schema sketch) - output format - example PM conversation that triggers it #### MCP code scaffold Generate a minimal TypeScript MCP server using @modelcontextprotocol/sdk. Single tool: [name, description, inputs]. Use stdio transport. Include package.json deps and a 1-line run instruction. Keep it under 60 lines. #### MCP use case brainstorm I'm a Growth PM at [company]. Brainstorm 5 MCP tools that would be highest-leverage for my daily work. For each: tool name | what it does | example prompt I'd type to use it | data source it touches. ### Exercise Design the Growth Workspace MCP - the 5 most valuable tools you'd give your AI. Use the code scaffold prompt to generate a TypeScript starter for the simplest one. Share with the colleague who built the existing MCP. ## Module 7: Proof of Concepts with AI URL: https://ai-course.dev/module/proof-of-concepts Duration: ~40 min · Category: Prototyping Build working demos before committing engineering resources. AI has collapsed the cost of PoCs dramatically - from Figma flows to website analysis to functional mini-apps that get stakeholder alignment fast. ### Lesson PoC as a PM superpower A clickable prototype kills 10 meetings. Tangible > slides for alignment. AI made the cost of "tangible" near-zero. Analyzing existing surfaces Drop a screenshot or URL into Claude. Ask for a CRO audit, copy improvements, A/B hypotheses. Compare to your instincts to find your unique signal. Functional demos in an hour HTML or React prototypes via Claude / v0 / Lovable. Aim for "convincing enough to provoke real reactions" - not pixel-perfect. Show don't tell Frame every PoC as a hypothesis, not a commitment. "Here's what it could look like - what does this make you want to change?" Section · Vibe Coding Vibe Coding - Building Without Writing Code Vibe coding is the practice of building functional software entirely through natural language - describing what you want, iterating on the output in plain English, and never directly writing or reading the underlying code. The term was coined by Andrej Karpathy in early 2025 and it's now a legitimate path to production-quality PoCs for non-engineers. For a PM, this is the highest-leverage skill shift of the last 2 years. It means "I'll prototype that" is no longer something only engineers can say. How it works in practice You describe the app or feature you want to build. The AI generates the code. You look at what it produced in the browser - not in the code editor - and describe what to change. You never need to understand the code to iterate on the output. The workflow: - Describe the full app in a single detailed prompt (what it does, who uses it, key interactions, visual style) - Review the output in the browser - does it do what you described? - Iterate in plain English: "the submit button does nothing", "add a step 3", "make the empty state more helpful" - Repeat until it matches your vision - Share the URL or screenshot for stakeholder feedback The right tool for the job Goal | Best tool | Hosted app with a real URL | Lovable | React component to hand off to an engineer | v0 | Complex local prototype, connecting to an API | Cursor | Quick single-file HTML demo, no setup | Claude Artifacts or ChatGPT Canvas | What makes a great vibe coding prompt The same rules as any prompt, but front-loaded with more product context: - Start with the user and their job-to-be-done, not the feature list - Describe the key screens and what happens on each - Specify visual style (dark/light, minimal/bold, desktop/mobile) - List what the app does not need to do (scope boundaries prevent over-engineering) - Include one example of a similar product for reference Bad: "Build me a dashboard" Good: "Build a single-page growth metrics dashboard for a B2B SaaS PM. It should show: weekly sign-ups (bar chart), activation rate (line chart), and a table of this week's experiments with status badges. Dark theme. Desktop only. No backend - use hardcoded sample data. Style reference: Linear.app" Where vibe coding breaks down - and how to handle it Vibe coding is not magic. It fails predictably in a few ways: - Context window overflow: Very large or complex apps lose coherence as the conversation gets long. Fix: start a fresh conversation with a summary of what's been built so far. - Cascading fixes: Fixing one thing breaks another. Fix: ask the AI to explain what it changed and why before accepting the output. - Hallucinated APIs: AI will sometimes write code that calls APIs or libraries that don't exist or work differently than expected. Fix: if something doesn't work, paste the error message back and ask AI to fix it with real, verified approaches. - Design drift: The visual output gradually drifts from your intent across many iterations. Fix: re-anchor with "go back to the original visual style I described at the start." The PM vibe coding mindset You are the product manager of the AI building your PoC. Your job is to give clear requirements, review output against those requirements, and give precise feedback when something is wrong. The same skills that make you a good PM - clarity of thought, precise feedback, knowing when something is right - make you a good vibe coder. What you are not doing: writing code, debugging logic, understanding architecture. What you are doing: product thinking at high speed. ### Prompts #### Website audit You are a senior CRO expert. Audit this page: [URL or paste HTML]. Return: 1. Top 3 friction points (with location on page) 2. 5 copy improvements (before → after) 3. 3 A/B test hypotheses ranked by expected impact 4. 1 structural change worth prototyping #### PoC builder brief Build a minimal working HTML prototype for a stakeholder demo. Idea: [describe in 2 lines]. Constraints: single file, inline CSS, no build step, mobile-friendly, realistic copy. Include 1 happy path + 1 edge case. Output the full HTML. #### Figma-to-insight Analyze this screen (image attached or described). Return: - The JTBD it serves - Likely drop-off points and why - The single first change you'd ship - What's missing for a great experience #### Vibe coding kickoff Build a [TYPE OF APP/PROTOTYPE] with the following spec: User: [WHO WILL USE THIS AND WHAT THEY'RE TRYING TO DO] Key screens: 1. [SCREEN 1 - what it shows and what actions are available] 2. [SCREEN 2 - what it shows and what actions are available] 3. [SCREEN 3 - if needed] Key interactions: - [INTERACTION 1] - [INTERACTION 2] Visual style: [dark/light theme, minimal/bold, desktop/mobile, style reference if any] Data: use realistic hardcoded sample data - no backend needed. Out of scope: [LIST WHAT IT DOES NOT NEED TO DO] Build this as a single self-contained file. Prioritise: correct interactions > visual polish > completeness. Make it look intentional, not like a wireframe. ### Exercise Build a PoC of a real growth idea at your company using the PoC builder prompt. Share with one colleague and capture the first 3 reactions verbatim. Run the website audit prompt on your company's site. Compare AI's output to your own instincts. What did you see that AI didn't? That's your edge. ## Module 8: AI Strategy for Growth PMs URL: https://ai-course.dev/module/ai-strategy-for-growth-pms Duration: ~35 min · Category: Strategy You've learned the tactics - now build the strategic layer: where AI genuinely creates leverage vs noise, how to build your personal AI stack, and how to bring your team along as a force multiplier. ### Lesson Leverage vs noise High-leverage AI = high-volume, repetitive, well-defined tasks where quality variance is costly. Low-leverage AI = creative one-offs you already love doing. The 4-layer personal AI stack - Brain - Claude / GPT for reasoning - Memory - Notion stores prompts, templates, learnings - Automation - n8n for scheduled flows - Integration - MCPs for persistent capability Bring the team along Run prompt clinics. Share templates in a single Notion. Apply growth-experiment mindset to AI adoption: small bets, fast feedback, double down on winners. Evals: how you know an AI workflow is actually good Most AI workflows are shipped on vibes. An eval is just a repeatable test set for a prompt or agent - the AI equivalent of a regression suite, and the single highest-leverage habit for anyone putting AI into a real process. - Collect 20 real inputs. Not synthetic ones - real tickets, real interview transcripts, real briefs, including the messy edge cases. - Write down what "good" looks like for each: the key facts that must appear, the format, what must never appear. - Run the prompt on all 20 and grade pass/fail. A 70% pass rate is fine for a draft assistant and unacceptable for anything customer-facing. - Change one thing at a time - prompt, context, or tier - and re-run. Keep the score in a sheet with the date. - Re-run when the model changes. Provider updates silently shift behaviour; your eval set is the only way you'll notice. For subjective outputs, use an LLM-as-judge: a second prompt that scores the output against your rubric. Spot-check the judge against your own grading on 10 cases before trusting it. Shipping AI features, not just using AI Using AI in your workflow and putting AI in your product are different jobs. If you own an AI feature, these are the decisions nobody else will make for you: Decision | What to get right | Cost & latency | Know your cost per request and p95 latency before launch. Route easy cases to the cheap tier; stream output so waiting feels shorter. | Failure UX | Design the unhappy path: low confidence, refusal, timeout. Never let a broken generation look like a confident answer. | Hallucination handling | Cite sources, show the retrieved snippet, and make it trivial to verify. "Trust me" is not a UI. | Human in the loop | Decide what AI drafts vs decides vs executes. Irreversible actions need a confirmation step. | Feedback loop | Thumbs plus a reason code, logged with the input. This becomes your eval set. | Guardrails | Input validation, output filtering, rate limits, and a kill switch you can flip without a deploy. | Rule of thumb for scoping: AI should reduce the cost of a step a user already takes. If the feature only works when the model is perfect, it isn't ready. The compounding PM Tools change every quarter. Taste and judgment for where AI creates leverage compounds for a career. Invest there. ### Prompts #### Personal AI stack audit Audit my current AI stack: [list tools, what you use each for]. Identify: - Gaps (what task has no tool) - Overlaps (where 2 tools do the same job) - The single highest-leverage addition for next 30 days, with rationale #### Team AI onboarding Design a 30-minute workshop introducing AI to a Growth team of 6 PMs. Include: - 2 specific demos (with the prompts I'll run live) - 1 hands-on exercise (10 min) - 1 reusable prompt template they'll keep - 1 follow-up ritual (weekly prompt share) #### AI opportunity map Map my recurring PM tasks to AI leverage. Tasks: [paste list]. Return a markdown table: | Task | Frequency | Leverage (L/M/H) | AI approach | Tool | Time saved / week | ### Exercise Run the AI opportunity map on your real weekly tasks. Pick the 3 highest-leverage. Write one concrete next step for each - that's your 90-day AI roadmap. Design a team workshop using the onboarding prompt. Then actually run it. Capture what landed and what didn't. ## Module 9: AI Toolbox - The Right Tool for Every Job URL: https://ai-course.dev/module/ai-toolbox Duration: ~40 min · Category: Tools The AI tool landscape moves fast and the choices are overwhelming. This module maps the tools that actually matter for product, growth, design, and engineering roles - with a deep dive on Cursor as the highest-leverage AI coding tool for non-engineers who want to build real PoCs. ### Lesson 1. The AI tool landscape (mapped by job to be done) Thinking & Writing Tool | Best for | Use it for | Claude (Anthropic) | PM, Growth, Designer | Research synthesis, long briefs, structured analysis, nuanced copy, complex instructions | ChatGPT (OpenAI) | All roles | Brainstorming, quick tasks, image analysis, breadth of capability | Gemini (Google) | All roles | Google Workspace tasks, real-time web search, data in Sheets | Perplexity | All roles | Research with live citations - replaces many search queries | Notion AI | PM, Growth, Designer | Synthesising notes, drafting docs, querying your existing knowledge base | Building & Coding Tool | Best for | Use it for | Cursor | Engineer, PM (PoCs) | AI-native code editor - pair programming, PoC building, codebase understanding | v0 by Vercel | Designer, PM | Generate React UI components from text descriptions - no coding required | Lovable | PM, Designer | Full-stack app generation from prompts - what this course was built with | Replit AI | Engineer, PM | Browser-based coding + AI - fast PoC deployment without local setup | GitHub Copilot | Engineer | In-editor code completion - the industry standard for engineers | Automation & Agents Tool | Best for | Use it for | n8n | PM, Growth, Engineer | Visual workflow automation - triggers, AI steps, tool connections | Make (Integromat) | PM, Growth | Similar to n8n - strong on SaaS integrations | Zapier | All roles | Simplest automation tool - lower ceiling but fast to set up | LangChain | Engineer | Framework for building multi-step AI agents in Python/JS | CrewAI | Engineer | Multi-agent orchestration - agents that collaborate on complex tasks | Design & Visual Tool | Best for | Use it for | Figma AI | Designer, PM | Design suggestions, copy generation, auto-layout within Figma | Canva AI | All roles | Quick visual creation, presentation design, social assets | Midjourney / DALL·E | Designer, PM | Image generation for mockups, concept visuals, presentations | Framer AI | Designer, PM | AI-generated websites and landing pages - design + publish in one tool | Research & Analysis Tool | Best for | Use it for | Perplexity | All roles | Research with cited sources - great for market research, competitor analysis | Elicit | PM, Growth | AI-powered literature and research synthesis | Rows AI | Growth, PM | AI-powered spreadsheets - query your data in plain English | PostHog AI | Growth, Engineer | AI-assisted product analytics | Meetings & Async Work This is the highest-frequency daily use case that most AI tool guides ignore. You spend 4–6 hours a day in meetings and async communication. AI can compress the overhead dramatically. Tool | Best for | Use it for | Granola | All roles | AI meeting notes - joins calls, transcribes, generates structured summaries automatically | Notion AI | PM, Growth, Designer | Summarise meeting notes, extract action items, draft follow-up comms from meeting context | Claude (with transcript) | All roles | Paste any transcript → extract decisions, actions, open questions, disagreements | Otter.ai | All roles | Real-time transcription + AI summary for any meeting | Fireflies.ai | All roles | Meeting intelligence - transcription, action items, CRM sync | The async work unlock. The highest-leverage application isn't meeting notes - it's async context compression. Long Slack threads, email chains, Notion docs that have grown unwieldy: paste them into Claude and ask "summarise this, extract open decisions, and list what's blocking progress." That's a 40-minute reading task turned into a 2-minute brief. For PMs specifically: end-of-sprint summaries, stakeholder update drafts, and weekly team digests are all tasks where AI removes the blank-page friction entirely. You paste the inputs (meeting notes, metrics, Slack threads), AI drafts the output, you edit and send. Granola - the recommended default for meetings. Granola runs in the background during any meeting (Zoom, Google Meet, Teams) and generates structured notes automatically. Unlike tools that record and transcribe only, Granola uses the transcript to generate a clean, formatted summary with action items. For growth PMs who are in 4–8 meetings a day: Granola means you stop taking notes entirely and start being fully present. The notes are there when you need them. The ROI is immediate. Agentic coding & terminal agents Tool | Best for | Use it for | Terminal coding agents (Claude Code, Codex CLI and equivalents) | PM, Engineer | Multi-file changes, repo-wide questions, long refactors that run while you review | IDE agents (Cursor, Windsurf, Copilot agent mode) | PM, Engineer | Building PoCs with a visual editor and inline diffs | Async coding agents (agents that open pull requests) | Engineer, PM | Small, well-specified tickets delegated end to end | The pattern to learn is delegation: write a crisp spec, let the agent work, then review a diff. The specific product you use will change; the review habit won't. Deep research & agentic browsing Tool | Best for | Use it for | Deep research modes (in ChatGPT, Claude, Gemini, Perplexity) | All roles | Multi-source market and competitor reports with citations, in 5–20 minutes | AI browsers & computer-use agents | PM, Growth | Filling forms, pulling data from tools without an API, repetitive click-work - always supervised | Treat every deep-research output as a first draft with footnotes: open at least three citations before you quote a number in a strategy doc. Spreadsheets, BI & voice Tool | Best for | Use it for | AI spreadsheets (AI formulas and enrichment inside Sheets/Excel-likes) | Growth, PM | Enriching lead lists, classifying rows, bulk summarising qualitative columns | BI copilots (natural-language layers on your warehouse) | PM, Growth | Self-serve metric questions without waiting on an analyst - verify the generated SQL | Realtime voice modes | All roles | Thinking out loud, interview practice, hands-free debriefs after a call | 3. Agentic coding in depth - Cursor as the worked example What is Cursor? Cursor is a code editor (built on VS Code) with AI deeply integrated throughout. Unlike GitHub Copilot which completes code as you type, Cursor lets you have full conversations with your codebase - you can ask it to build features, explain code, fix bugs, and generate entire applications from a description. For a non-engineer, Cursor is the tool that makes "I'll build a quick prototype" actually possible - without needing to understand every line of code you're working with. Why Cursor vs v0 or Lovable? | Cursor | v0 | Lovable | Best for | Full apps, complex PoCs, working with existing code | React components, UI snippets | Full-stack apps from scratch | Requires | Basic comfort with files and a terminal | None | None | Output | Runs locally, fully customisable | Component code to paste | Hosted app | AI depth | Deepest - full codebase understanding | UI-focused | App-focused | Learning curve | Medium | Low | Low | Use Cursor when: you want to build something more complex than a UI component but don't want to fully hand over control to an app generator. Great for: data dashboards, multi-step flows, PoCs that connect to a real API. Getting started with Cursor (non-engineer path) - Step 1 - Install Cursor. Download from cursor.com. It looks like VS Code. If you've never opened a code editor, that's fine - you'll be using the AI chat panel more than the code itself. - Step 2 - Start a new project with AI. Open Cursor, press Cmd+Shift+P → "New File". Then open the AI chat panel (Cmd+L). Type: "Create a simple HTML and JavaScript app that [describe what you want to build]. Make it self-contained in one file. Use Tailwind CSS from a CDN." Cursor will generate the full file. - Step 3 - Iterate in plain English. Don't edit the code manually. Instead, select the part that's wrong, press Cmd+K, and tell Cursor what to change: "Make this button green", "Add a second step to this flow", "The table isn't showing - fix it." Cursor edits the code based on your instruction. - Step 4 - Preview your PoC. Open the generated HTML file in a browser. That's your prototype. For more complex apps, Cursor can run a local server - ask it: "How do I preview this?" and follow the instructions. - Step 5 - Explain code you didn't write. If you receive a codebase from an engineer, open it in Cursor and ask: "Explain what this file does", "Where does the sign-up flow start?", "What would break if I changed this value?" - Cursor answers in plain English. Practical PoC examples for PMs with Cursor - "Build a simple A/B test calculator - inputs: control conversion rate, variant conversion rate, sample size. Output: statistical significance and recommendation." - "Build a one-page onboarding flow mockup with 3 steps, a progress bar, and a completion screen." - "Build a dashboard that reads from this CSV and shows a bar chart of sign-ups by week." - "Build a prototype of a referral page with an invite input, a copy-link button, and a counter showing how many friends have been invited." What Cursor can't do for non-engineers - Deploy to a real URL without some engineering help (use Lovable or v0 for that) - Connect to your production database (use it for PoCs, not production code) - Replace a real engineer for anything that ships to users The PM Cursor workflow in one sentence: describe what you want to build in plain English → iterate in plain English → open in browser → show stakeholders. 4. Choosing your personal AI stack Your role-specific stack recommendation appears below ↓ ### Prompts #### Build a Cursor PoC from a brief I want to build a quick prototype in Cursor. Here is what I need: Goal: [DESCRIBE WHAT THE PROTOTYPE SHOULD DO] Users: [WHO WILL SEE THIS DEMO] Key interactions: [LIST 2–3 THINGS THE USER SHOULD BE ABLE TO DO] Visual style: [DESCRIBE - e.g. clean, dark theme, minimal] Build this as a single self-contained HTML file using vanilla JavaScript and Tailwind CSS from CDN. Make it look professional and polished. Prioritise the core interaction over completeness. Add placeholder data where needed. Include clear comments so I can understand what each section does. #### Which tool should I use for this task? I need to: [DESCRIBE THE TASK IN DETAIL] My role is: [ROLE] My technical comfort level is: [non-technical / some experience / comfortable with code] I need the output to be: [a shareable prototype / a local demo / production-ready / a one-off analysis] Timeline: [how long I have] Recommend the best AI tool for this specific task. Include: which tool, which tier/model, how to approach it, and any pitfalls to avoid. Be specific. #### Audit my AI stack Here is my current AI stack: [LIST YOUR TOOLS] My role is [ROLE] at [COMPANY]. My most common tasks are: [LIST 5–8 TASKS]. Audit my stack: 1. Which tools am I under-using? 2. What important use cases have no good tool coverage? 3. Where am I using the wrong tool for a task? 4. What one addition would have the highest leverage? 5. What, if anything, should I drop? Be specific and direct. Base your answer on my actual tasks, not generic advice. #### Meeting transcript processor Here is the transcript from a [TYPE - team meeting / stakeholder review / user interview / sprint planning]: [PASTE TRANSCRIPT] Extract and structure the following: 1. Key decisions made (with who made them) 2. Action items (owner + deadline if mentioned) 3. Open questions that still need answers 4. Key disagreements or tensions surfaced 5. One-paragraph summary for someone who wasn't there Return as structured markdown. Be specific - use names and direct quotes where relevant. #### Async thread summariser Here is a long [Slack thread / email chain / Notion comment thread]: [PASTE CONTENT] Give me: 1. What this thread is about (1 sentence) 2. The current state / what's been decided 3. What's still unresolved or being debated 4. Who I need to follow up with and why 5. Suggested next action (what should happen to move this forward) Be direct. If nothing is unresolved, say so. ### Exercise 🛠 Exercise 1 - Build your first Cursor PoC. Install Cursor (cursor.com). Pick a growth idea or feature you've been wanting to prototype. Use the "Build a Cursor PoC from a brief" prompt to generate a working HTML file. Open it in a browser. Share it with one colleague. Time yourself - the goal is under 45 minutes from idea to demo. 🎯 Exercise 2 - Audit your current AI stack. Use the "Audit my AI stack" prompt with your honest current tool list. Identify the one gap that would have the highest leverage. Spend 20 minutes exploring the tool that fills it. 💡 Exercise 3 - Map a task to the right tool. Pick 5 tasks you've done manually in the last week that felt slow or repetitive. For each one, use the "Which tool should I use?" prompt. You'll likely find 2–3 tasks that have a much better tool than what you're currently using. ## Module 10: AI Best Practices - Do's, Don'ts & Mental Models URL: https://ai-course.dev/module/ai-best-practices Duration: ~30 min · Category: Reference The meta-skill layer. This module teaches not just how to use AI, but how to think about using it well - the difference between a PM who gets 10% productivity gains and one who gets 10x leverage. A permanent reference you'll come back to. ### Lesson ### Prompts #### Build my AI context doc You are helping me set up my AI working context. Ask me the following questions one at a time, then compile my answers into a reusable context document I can paste at the start of AI conversations: 1. What is your role and company? 2. Who is your ICP (ideal customer profile)? 3. What are your current top 3 priorities or OKRs? 4. What are 3 words that describe your brand voice? 5. What are 2 examples of output you've been proud of recently? 6. What should AI never do when helping you? Compile into a structured "AI Context" document in markdown. #### Audit my prompt Here is a prompt I'm currently using: [PASTE YOUR PROMPT] Audit it against these 5 criteria: 1. Role clarity - is there a clear role framing? 2. Context richness - does it have enough background? 3. Output format - is the desired format specified? 4. Constraint clarity - are constraints (length, tone, what to avoid) clear? 5. Examples - would a few-shot example improve this? For each criterion: score (1–5), what's missing, and a specific improvement. Then rewrite the full prompt with all improvements applied. #### Design my prompt library I am a [ROLE] at [COMPANY]. My most common AI use cases are: [LIST 5–8 TASKS]. Help me design a prompt library structure for Notion. For each use case: 1. Prompt template name 2. When to use it 3. Starter template with [PLACEHOLDERS] 4. Tips for getting the best output Return as a structured Notion-ready markdown doc. ### Exercise 🛠 Exercise 1 - Build your AI context doc. Use the "Build my AI context doc" prompt to create your personal context document. Save it somewhere accessible (Notion, Apple Notes, a pinned Claude conversation). Use it in your next 5 AI conversations and notice the difference. 🎯 Exercise 2 - Audit your 3 most-used prompts. Pick 3 prompts you use regularly. Run each through the "Audit my prompt" template. Rewrite them based on the feedback. These 3 improved prompts are the seed of your prompt library. 💡 Exercise 3 - Apply the 80/20 rule to your week. List your top 10 recurring tasks. For each, mark: AI leverage (high/medium/low) and whether it's in the 80% (drafting, structuring, synthesising) or 20% (strategic judgment, relationships, decisions). This map tells you exactly where to invest AI time. ## Module 11: AI Ethics & Responsible Use for PMs URL: https://ai-course.dev/module/ai-ethics Duration: ~30 min · Category: Ethics This module is opinionated. Not philosophical - practical. As a PM, you make decisions about when and how AI touches your users, your team, and your company's data. This module gives you a clear framework for where to draw lines, what risks are real vs overhyped, and how to be a PM who uses AI responsibly without being paralysed by it. ### Lesson Section 1 - The three risks PMs actually need to care about There are a lot of AI ethics concerns in the world. Most of them are not your problem as a PM using AI for growth and product work. Three of them are. Risk 1: Data privacy - what you're allowed to paste Every time you paste content into Claude, ChatGPT, or any AI tool, you are sending that data to a third-party server. For most day-to-day work this is fine. It is not fine when that data includes: - Individual user data (names, emails, behaviour tied to identifiable people) - Financial data covered by confidentiality agreements - Unreleased product roadmaps or M&A information under NDA - Employee performance data or HR records - Any data covered by GDPR, HIPAA, or similar regulation The test: would you be comfortable if this data appeared in a legal discovery? If not, don't paste it. What you should do instead: - Anonymise data before pasting (replace names/emails with "User A", "Company B") - Use enterprise tiers with data processing agreements (Claude for Enterprise, ChatGPT Enterprise) for sensitive work - Build local automations (n8n on-premise, self-hosted models) for workflows that touch sensitive data - Know your company's AI data policy - if one doesn't exist, push for one Risk 2: Hallucination at scale - automations that lie confidently AI hallucination (generating plausible but false output) is a manageable risk in one-off use. It becomes a serious risk when you automate. An automation that generates incorrect product descriptions, wrong pricing, or fabricated statistics and publishes them without review is a trust and legal liability. The rule: human review checkpoints are mandatory for any AI output that goes directly to users or external stakeholders. This means: - Don't automate customer-facing copy without a review step - Don't automate data summaries to leadership without a sanity-check step - Don't automate anything that would be embarrassing or harmful if wrong - without a gate You are not being overly cautious. You are being a good PM. Risk 3: Bias in, bias out AI models reflect the data they were trained on. That data contains human biases - around gender, ethnicity, geography, economic background. When you use AI to generate copy, analyse user feedback, or make prioritisation recommendations, those biases can slip into your outputs silently. Practical implications: - User research synthesis: AI may underweight feedback from underrepresented user segments if they appear less frequently in the training data. Always sanity-check "whose voice is missing here?" - Copy generation: Review AI-generated copy for language that implicitly assumes a particular user demographic - Prioritisation: If you're using AI to rank feature requests or hypotheses, audit the methodology - AI rankings are not neutral You can't eliminate bias. You can build review practices that catch it. Section 2 - Strong opinions on what PMs should and shouldn't do These are recommendations, not rules. But they're based on real failure modes - not theoretical ones. You should: - ✅ Use enterprise AI tiers for anything that touches company data. The free tiers of most AI tools train on your inputs by default. Enterprise tiers don't. If you're pasting internal strategy docs, user research, or financials - this matters. - ✅ Be transparent with your team when AI contributed to work. Not "I used AI to write this so it doesn't count" - but "here's how I used AI, here's where I applied my judgment." Normalise it. Hiding it creates a culture of AI shame that slows adoption. - ✅ Own the output, always. If you submit an AI-generated experiment brief, it's your brief. If it's wrong, that's on you. AI is a tool - you're the PM. - ✅ Tell users when they're interacting with AI. If you're building AI-powered features that interact with users, clear disclosure is both ethical and increasingly legally required in many markets. - ✅ Audit your automations for quality drift. AI models update. What produced good output in January may produce worse output in June. Check automations quarterly. You shouldn't: - ❌ Don't paste identifiable user data into consumer AI tools. This is not theoretical - it's a GDPR violation. Anonymise first, always. - ❌ Don't use AI to make decisions about people. Hiring decisions, performance reviews, user bans - these require human judgment and carry legal weight. AI can inform, not decide. - ❌ Don't ship AI-generated copy without reading it. Every word. Not because AI is usually wrong - because the one time it's wrong, it's your product that said it. - ❌ Don't treat AI confidence as accuracy. AI sounds equally confident when it's right and when it's hallucinating. Confidence is not a signal of correctness. Verify anything that matters. - ❌ Don't build AI features that users can't opt out of. Especially personalisation, automated communications, and anything that uses user data to make decisions about them. Give people a way out. Section 3 - AI governance in practice (EU AI Act & friends) Principles are no longer the whole story: AI use is now regulated. You don't need to be a lawyer, but you do need to know which bucket your use case falls into. - Risk tiers. The EU AI Act sorts systems into prohibited (e.g. social scoring, most emotion recognition at work), high-risk (hiring, credit, education, essential services), limited-risk (transparency duties), and minimal-risk. Most internal PM workflows are minimal-risk - hiring and performance use cases are not. - Transparency duties. Users must be told when they're interacting with an AI system, and synthetic media should be labelled. Bake this into copy, not into a footnote. - AI literacy. Organisations deploying AI are expected to ensure staff are trained on its use and limits - one reason a course like this belongs in onboarding. - Data residency & vendor terms. Check where the provider processes and stores data, whether your inputs train their models, and what the retention window is. Enterprise/API tiers usually differ sharply from consumer plans. - Keep an inventory. A simple sheet - use case, tool, data types, owner, risk tier, human-review step - answers 90% of what legal and security will ask. If a use case touches hiring, credit, health, children, or biometric data, involve legal before you prototype, not after. Section 4 - The GDPR practical checklist for PMs using AI This is not legal advice. It is a practical checklist for avoiding the most common mistakes. Before pasting data into any AI tool, ask: - ☐ Does this data contain names, emails, or identifiers tied to real people? → Anonymise first - ☐ Is this data covered by an NDA or confidentiality agreement? → Check before pasting - ☐ Does this tool have a data processing agreement (DPA) with my company? → Check with your legal/compliance team - ☐ Is this tool's enterprise tier in use, or is data being used for training? → Know your contract - ☐ If this data were leaked, would it harm a user, employee, or the company? → If yes, don't paste it Before shipping an AI-powered feature: - ☐ Is there a disclosure to users that AI is involved? - ☐ Is there a human review step for AI outputs that affect users directly? - ☐ Is there an opt-out mechanism for users who don't want AI-generated content or decisions? - ☐ Have you audited the AI outputs for bias in a representative sample? - ☐ Do you have a process for monitoring quality drift over time? ### Prompts #### AI feature ethics audit I am planning to build an AI-powered feature: [DESCRIBE FEATURE]. It will: [WHAT IT DOES] It will use: [WHAT DATA IT USES] Users will: [HOW USERS INTERACT WITH IT] Audit this feature for: 1. Data privacy risks (what data is being processed, by whom, under what terms) 2. Hallucination risks (what happens if the AI output is wrong) 3. Bias risks (whose experience might be negatively affected) 4. Transparency requirements (do users know AI is involved) 5. Legal considerations (GDPR, consumer protection, sector-specific regulation) For each risk: severity (high/medium/low), likelihood, and a specific mitigation recommendation. Be direct - I want to know the real risks, not reassurance. #### Data anonymisation helper I need to paste the following data into an AI tool for analysis, but it contains personally identifiable information. Anonymise it for me: [PASTE DATA] Rules: - Replace all names with "User [A/B/C...]" or "Company [1/2/3...]" - Replace all email addresses with "user@example.com" - Replace all specific dates with relative dates ("3 months ago", "last Tuesday") - Replace all specific revenue figures with ranges ("€10K–50K") - Keep all other context intact - the goal is to preserve analytical value while removing identifiers Return the anonymised version only. #### AI policy draft for your team Help me draft a simple AI usage policy for my product/growth team. We are a [SIZE] team at a [TYPE] company. Our main AI use cases are: [LIST]. The policy should cover: 1. What data can and cannot be pasted into AI tools 2. Which tools are approved for which use cases 3. Human review requirements for AI-generated content 4. Disclosure standards (when to tell colleagues / users that AI was used) 5. How to handle AI errors or quality issues Keep it practical and under 500 words. Write it as a living document we can update, not a legal contract. Opinionated recommendations are better than vague guidelines. ### Exercise 🛠 Exercise 1 - Audit your current AI usage. Go through your last 5 AI interactions. For each one: what data did you paste? Was any of it personally identifiable? Was it covered by confidentiality? This is not to make you feel bad - it's to make the risk concrete. One finding from this audit is worth 10 pages of theory. 🎯 Exercise 2 - Run the GDPR checklist on a real automation. Take your most data-rich existing automation (insight report, NPS synthesis, translation agent). Run it through the GDPR practical checklist above. What passes? What needs a fix? Document the gaps and prioritise the highest-risk one. 💡 Exercise 3 - Draft your team's AI policy. Use the "AI policy draft" prompt to generate a first draft for your team. Share it with your manager or a colleague and ask: "Is there anything here we're currently doing wrong?" One honest answer to that question is worth more than the whole policy. ## Glossary ### Large language model (LLM) URL: https://ai-course.dev/glossary/llm A model trained on vast amounts of text that generates language by predicting the most likely continuation of its input. A large language model is a pattern-completion system, not a database or a search engine. It has learned statistical associations across billions of examples, which is why it can write fluently about almost anything - and why it can be confidently wrong. For product people, three properties matter: it has no memory between conversations unless you give it one, its output quality tracks the quality of the context you provide, and it optimises for plausible rather than true. ### Token URL: https://ai-course.dev/glossary/token The unit an LLM reads and writes - roughly three quarters of a word in English. Models don't see letters or words, they see tokens. Pricing, rate limits and context windows are all measured in tokens, so a rough conversion is useful: 1,000 tokens ≈ 750 words ≈ 1.5 pages. If a workflow feels expensive or slow, count tokens before blaming the model: pasting a 60-page PDF into every request is usually the real cost driver. ### Context window URL: https://ai-course.dev/glossary/context-window The maximum amount of text a model can hold in mind at once, including your prompt, the history and any attached documents. Everything the model can 'see' when answering has to fit in the context window: system prompt, conversation history, retrieved documents and the current question. Bigger is not automatically better. Answers measurably degrade when a long window is filled with loosely relevant material, so curating what goes in beats maximising how much goes in. ### Prompt engineering URL: https://ai-course.dev/glossary/prompt-engineering Designing the instructions you give a model so it reliably produces the output you need. A strong prompt usually has five layers: role, context, task, constraints and examples. The single highest-return habit is stating the output format explicitly at the end of the prompt. Prompt engineering is a repeatable craft, not a bag of magic words. If a prompt only works with one specific model version, it is fragile. ### Context engineering URL: https://ai-course.dev/glossary/context-engineering Designing everything the model can see besides your prompt: files, retrieved documents, memory, tool outputs and history. As models got better at following instructions, the bottleneck moved from wording to context. Context engineering is the discipline of assembling the right material, in the right structure, at the right moment. Practical rules: curate rather than dump, connect changing sources instead of pasting stale copies, decide deliberately what persists across sessions, and restate the goal in long threads before context rot sets in. ### RAG (retrieval-augmented generation) URL: https://ai-course.dev/glossary/rag A pattern where relevant documents are retrieved from your own data and inserted into the prompt before the model answers. RAG is how you make a general model answer questions about your specific content - a help centre, a research repository, a product spec library - without retraining anything. The quality of a RAG system is mostly the quality of its retrieval. If the wrong chunks are fetched, no amount of prompt tuning fixes the answer. ### MCP (Model Context Protocol) URL: https://ai-course.dev/glossary/mcp An open standard that lets any AI client discover and call your tools and data sources through one interface. With a direct API integration, you decide when the tool gets called. With MCP, the model decides - it discovers the available tools, reads their descriptions and calls them when relevant. You write the server once and any MCP-compatible client can use it, which is why MCP has become the default way to connect AI assistants to internal systems. ### AI agent URL: https://ai-course.dev/glossary/ai-agent A system where a model plans a sequence of steps and calls tools to accomplish a goal, rather than just answering once. The useful mental model is a loop: perceive, plan, act with tools, then check the result and repeat. Anything that only produces text in one shot is an assistant, not an agent. Agents earn their complexity when the task is repetitive, multi-step and tolerant of a review step. Irreversible actions should always keep a human in the loop. ### Hallucination URL: https://ai-course.dev/glossary/hallucination Output that is fluent and plausible but factually wrong or invented. Hallucination is not a bug that will be patched out; it is a consequence of models optimising for likely continuations. It becomes dangerous when it is automated and nobody reads the output. Mitigations that work in practice: ground answers in retrieved sources, ask for citations, require the model to say when it does not know, and design a UI that makes verification cheap. ### Eval URL: https://ai-course.dev/glossary/eval A repeatable test set that scores an AI workflow against examples of what 'good' looks like. An eval is the AI equivalent of a regression suite: 20 real inputs, a written definition of a correct answer, a pass rate, and a date. Without one, every prompt change is a guess. Re-run your evals whenever the model, the prompt or the context changes. Provider updates shift behaviour silently. ### LLM-as-judge URL: https://ai-course.dev/glossary/llm-as-judge Using a second model call to grade the output of the first against a written rubric. This is how teams evaluate subjective output - tone, helpfulness, faithfulness - at a scale where human grading is impossible. Always calibrate the judge: grade ten cases yourself and check the judge agrees before trusting its scores. ### System prompt URL: https://ai-course.dev/glossary/system-prompt Standing instructions that apply to every message in a conversation or project. The system prompt is where role, audience, tone, definitions and hard constraints belong. Put anything you would otherwise retype into it. In Claude Projects, Custom GPTs and Gems, the system prompt is what turns a blank chat into a pre-briefed colleague. ### Chain-of-thought URL: https://ai-course.dev/glossary/chain-of-thought Asking a model to reason step by step before answering, which improves accuracy on multi-step problems. Explicit reasoning gives the model room to decompose a problem instead of guessing an answer in one leap. Reasoning tiers do this internally by default. For PM work, the practical version is: 'List the assumptions, then evaluate each, then give your recommendation.' ### Temperature URL: https://ai-course.dev/glossary/temperature A setting that controls how random a model's output is - low for consistency, high for variety. Low temperature suits extraction, classification and anything you will parse. Higher temperature suits brainstorming, naming and copy variants. If an automation produces inconsistent results run to run, temperature is the first dial to check. ### Embedding URL: https://ai-course.dev/glossary/embedding A numeric representation of text that lets software compare meaning rather than keywords. Embeddings power semantic search, deduplication and clustering - for example grouping hundreds of support tickets by theme without predefined tags. They are cheap and fast compared with generation, which makes them ideal for the retrieval half of a RAG system. ### Fine-tuning URL: https://ai-course.dev/glossary/fine-tuning Further training a model on your own examples so it internalises a format or style. Fine-tuning is for consistent behaviour, not for teaching facts. If you need the model to know something, retrieve it instead. For most product teams, better prompts and better context beat fine-tuning on cost, speed and maintenance. ### Multimodal URL: https://ai-course.dev/glossary/multimodal A model that handles more than text - images, audio, video or screenshots - in the same conversation. Multimodal input is quietly one of the most useful PM features: paste a dashboard screenshot, a competitor's pricing page or a Figma export and ask for analysis. Treat visual interpretation like any other output: verify the numbers it reads back to you. ### Vibe coding URL: https://ai-course.dev/glossary/vibe-coding Building working software by describing what you want in natural language and iterating on what the AI produces. Vibe coding lets a non-engineer turn an idea into a clickable, functional prototype in an afternoon - which changes how product decisions get made. It breaks down on production concerns: security, scale, data models and edge cases still need engineers. ### Guardrails URL: https://ai-course.dev/glossary/guardrails The checks around a model that constrain what can go in and what can come out. Guardrails include input validation, output filtering, allow-lists of tools, rate limits, confidence thresholds and a kill switch you can flip without a deploy. Design them before launch. Retrofitting guardrails after an incident is far more expensive than building them in. ### EU AI Act URL: https://ai-course.dev/glossary/eu-ai-act European regulation that classifies AI systems by risk and attaches obligations to each tier. Systems are sorted into prohibited, high-risk, limited-risk and minimal-risk. Most internal productivity use is minimal-risk; hiring, credit, education and essential services are not. Two duties catch product teams by surprise: telling users when they are interacting with AI, and ensuring staff have adequate AI literacy for the systems they deploy.