AI Guides › Playbooks
By Nigel Guy · 6 min read
Most people learning AI do the first two stages, feel fluent, and quietly stop. They can chat with an assistant and write a decent prompt, so it feels like the job is done. What they have is vocabulary. The skills that change your work, building something small and then making it run without you, sit further down the road and are where nearly everyone drops off.
The rule: do not move to the next stage until you can pass its exit check with something you made, not something you watched.
This is JulieMango's own roadmap, not a recognised curriculum, and nobody certifies it. Course facts below were checked against the providers' own pages on 2026-10-04; where a page did not state something (usually price), we say so.
| Stage | Question it answers | Free course to start with | Exit check |
|---|---|---|---|
| 1. Language | How do I talk to AI well? | Claude Academy: AI Fluency: Framework and Foundations | You can brief an assistant with goal, context and format, and say what you checked afterwards |
| 2. Machine | What is it actually doing, and where does it fail? | Claude Academy: AI Capabilities and Limitations | You can predict, before asking, which of your tasks it will get wrong |
| 3. Systems | How do I stop starting from zero each time? | OpenAI Academy pathways, or Anthropic's product courses | Your three most repeated tasks each have a saved setup |
| 4. Build | Can I make a small thing that works? | Kaggle Learn Python, then Hugging Face Agents Course | One working script or agent, run by you, on real input |
| 5. Automate | Can it work while I am not there? | Anthropic's MCP and agent-skills courses | One task runs on a trigger, and you review the output, not the process |
Stages one to three are comfortable. You stay in a chat window, nothing breaks, and every session feels productive. Stage four asks you to run code, read an error message and fix it. That is the point where the novelty stops paying you back, so people decide they are "not technical" and return to chatting.
The fix is to make stage four small. You are not training a model or becoming an engineer. You are making one thing that takes an input and produces an output, once, on your own data.
Claude Academy lists "AI Fluency: Framework and Foundations" as a free course of 14 lessons and a quiz, about 4 hours. It teaches a framework named the 4D framework: Delegation, Description, Discernment and Diligence. Take it as a structure for the habit, not as a qualification; the page does not mention a certificate.
Practise with a brief you reuse:
You are a patient work colleague helping me with [TASK].
Goal: [WHAT_A_GOOD_RESULT_LOOKS_LIKE]
Context: [WHO_IT_IS_FOR_AND_WHAT_YOU_ALREADY_KNOW]
Material: [PASTE_YOUR_SOURCE_TEXT]
Format: [LENGTH_AND_STRUCTURE]
Rules:
- Use only my material for facts. If you need something I have not given you, ask me first.
- Do not invent names, numbers or quotes.
Before answering, check your draft against the goal and rules.
Then give me the draft, and a list of every factual claim tagged "from my material" or "not from my material".
Fill in the square-bracket parts. Exit check: you can say what you verified in your last three outputs.
"AI Capabilities and Limitations" is listed on Claude Academy as free, 13 lessons and a quiz, about 3.5 hours, covering next-token prediction, knowledge, working memory, steerability and context limits. This is the stage that makes you sceptical in a useful way. After it, write down five tasks and predict where the assistant will fail before you test it. Being wrong about your own prediction is the point.
Move from one-off chats to saved setups: a reusable brief, a project, a standing instruction file. OpenAI Academy describes pathways such as "Apply AI at Work" and "Build with AI", plus events and communities; its page does not state a price, so check before assuming it is free. Anthropic's catalogue includes introductory courses such as Claude 101 and Introduction to Claude Cowork; the catalogue page does not state cost either. Whichever you choose, the deliverable is the same: your three repeated tasks, each with a saved setup.
Kaggle Learn lists Python, Data Visualization and Pandas tutorials; we could not confirm its free status or certificates from the page text, so check it. For building, the Hugging Face Agents Course states it is entirely free, including certification, and needs basic Python knowledge. It runs from agent fundamentals through frameworks (smolagents and LangGraph) to use cases and a final assignment. It has a fundamentals certificate (Unit 1 only) and a full one, with no deadline.
Use this prompt to get a first project sized correctly:
You are a coding tutor for a beginner who knows [YOUR_PYTHON_LEVEL].
My goal: a small script that takes [INPUT_TYPE_I_HAVE] and produces [OUTPUT_I_WANT].
Sample input I can share: [PASTE_ONE_REAL_EXAMPLE]
Do this in order:
1. Restate the task in two sentences and ask me any questions you need answered. Wait for my answers.
2. Propose the smallest version that could work, under 40 lines.
3. Explain each part in plain English.
4. Tell me exactly how to run it and what I should see.
Do not add features I did not ask for. Do not assume libraries are installed; tell me how to install them.
Before answering, check that every step can be followed by someone who has never run a script.
Exit check: it runs on real input, and you can explain what each part does.
Anthropic's catalogue lists "Introduction to Model Context Protocol", "Introduction to agent skills" and "Introduction to subagents". Third-party summaries describe the MCP course as building servers and clients in Python, but we could not verify that on the catalogue page itself. The principle matters more than the course: pick one weekly task, write it as ordered steps, let it run on a trigger, and put yourself in the last step as reviewer.
We will not give you a number of weeks. Course lengths are listed (4 and 3.5 hours for the two Claude Academy courses), but your pace, your Python starting point and how much real work you apply it to decide the rest. Measure progress by exit checks passed, not calendar time.