Most prompt engineering courses charge a few hundred pounds to teach you something you could reasonably learn in a weekend. The free ones are often just as good, and sometimes better, because nobody is padding out the syllabus to justify a price tag.
The catch isn’t cost. It’s knowing which of the hundreds of free options is a genuine curriculum and which is a twelve-page PDF that ends in a form asking for your work email.
What a decent free prompt engineering course actually teaches
Strip away the branding and almost every useful course covers the same handful of skills. If a syllabus doesn’t mention at least six of these, it’s a brochure, not a course:
- Role and context setting, and why a vague persona like “you are an expert” does almost nothing
- Few-shot examples and how many you actually need before the model stops improving
- Structured output, usually JSON or a fixed table, plus what to do when the model adds a chatty preamble anyway
- Chain-of-thought and decomposition, including when step-by-step reasoning makes results worse
- Evaluation, meaning how you test 20 prompts against each other instead of eyeballing one output
- Prompt chaining, where one call extracts and the next one rewrites
- Failure modes: hallucinated citations, drift over long conversations, overfitting to your examples
The evaluation module is the tell. Courses that skip it teach you to write prompts by vibes, which is why people finish them and still can’t explain why one version of a prompt beats another.
Four free options that hold up
DeepLearning.AI’s short course with OpenAI
Roughly ninety minutes across nine short lessons, taught by Isa Fulford and Andrew Ng. It leans on a Jupyter notebook, so you’re editing prompts and rerunning cells rather than watching someone else do it. The section on building a custom chatbot is dated in places but the fundamentals still land.
Learn Prompting
Open source, browser-based, and unusually thorough for something with no paywall. The basics section takes an afternoon; the advanced material on retrieval and prompt injection is where it earns its reputation. There’s a reason it keeps showing up in discussions of free generative AI courses worth your evenings.
Anthropic’s prompt engineering tutorial
Nine chapters, run through a spreadsheet rather than a code editor, which sounds odd until you realise it lowers the barrier for non-programmers. The exercises on system prompts and prefill are the clearest explanation of those mechanics I’ve seen in a free resource.
Microsoft’s Generative AI for Beginners
Twenty-one lessons on GitHub covering prompting alongside RAG, fine-tuning and responsible AI. It’s broader than prompt engineering alone, so treat it as the next step rather than a starting point.
How to spot a lead magnet dressed up as a course
I’ve signed up to maybe fifteen of these. The patterns are consistent.
- No practice environment, no downloadable examples, no way to test anything
- A call-to-action at the end of every single module
- Zero discussion of prompts that fail
- A certificate for twenty minutes of video
- Last updated in early 2023 and never touched since
One more test: search the course page for the word “eval”. If it doesn’t appear, the instructor probably hasn’t shipped anything with prompts in production. The same filters apply when you’re sizing up paid options, which is the core argument in how to tell good AI courses from expensive ones.
A schedule that gets you to the end
Completion rates on free courses are grim, mostly because there’s no deadline and no sunk cost. Twenty minutes a day for two weeks beats a heroic six-hour Saturday every time.
Watch one module in the morning. In the afternoon, take a task you already do at work, a summary, a client email, a data cleanup, and write the prompt from scratch without looking at the course. Then compare. Keep both versions in a single file called something like prompt-log.md, with a one-line note on what changed and whether it helped.
By day ten you’ll have around twenty logged prompts, half of which will be genuinely reusable. That file is worth more than the certificate. If you need help building the habit, there’s practical advice in free AI classes online and how to actually finish one.
Practise on your own mess, not the course’s toy examples
Every course uses the same sanitised examples: summarise this paragraph about photosynthesis, extract dates from a press release. They’re clean, short, and nothing like your actual inputs.
Swap them out early. Take a rambling three-paragraph customer email and ask for a two-sentence summary plus a suggested reply. Feed in a messy spreadsheet column pasted as plain text and ask for a normalised list. The friction you hit doing that is the real curriculum, because that’s where you discover the model needs an explicit output format, or that your examples contradicted each other.
Do this for a week and you’ll write better prompts than someone who watched all nine lessons passively. It’s the same gap that separates useful training from expensive training, a point worth keeping in mind once you start looking at AI courses for professionals.
Building a prompt library you’ll still use in six months
The last thing worth doing, and the one almost nobody does, is turning your log into a small library. Aim for twelve to fifteen prompts, each with a header comment explaining the task, the model it was tested on, and the date.
Store the version that worked, not the version that was elegant. Add a line at the top of each with the failure you were fixing, for example “model kept inventing ticket IDs, so I added a rule denying it access to the ID field”. Six months from now, that note will save you an hour of confused debugging.
Then test the library against a model upgrade before you trust it. Prompts that leaned on a quirk of one version often break quietly on the next, and the only defence is a handful of known inputs with expected outputs you can rerun in five minutes. That habit, more than any certificate, is what separates people who use these tools well from people who keep rewriting the same prompt and hoping.