There is a pattern with vendor learning material where the quality of the reference documentation and the quality of the teaching move in opposite directions, and Google's Gemini developer resources are a clean example. As reference, this is good. As a place to learn, it is a pile of parts with no instructions. Take the good first.
The quickstart genuinely is quick. Get an API key from AI Studio, make a call, see a response, and you are running in well under an hour. The free tier is generous enough that you can build and iterate on something real before any billing conversation happens, which lowers the barrier more than people give it credit for. Anyone who has tried to prototype against an enterprise API with a procurement process in front of it will appreciate what that is worth.
AI Studio itself is a sensible piece of design. Being able to test a prompt in a browser, adjust parameters, see how the model behaves with different system instructions, and only then move to code is the correct order of operations, and it saves a great deal of time compared with iterating through a script. It also makes the model's behaviour visible in a way that reading documentation does not, which matters because a lot of what you need to learn about these systems is empirical rather than conceptual. The cookbook is the best thing here by a distance.
It is a repository of runnable notebooks covering the actual capabilities, function calling, structured output, multimodal input, long context handling, embeddings, agent patterns, and the examples are current and work. If somebody asked me how to learn to build on Gemini, my answer would be to clone the cookbook and work through the notebooks relevant to what you are building, ignoring almost everything else. That is where the practical knowledge lives. The documentation on specific capabilities is also more detailed than I expected.
Function calling in particular is covered properly, including the awkward parts around schema definition and handling model decisions to call nothing. Structured output gets real treatment. There is decent coverage of building with third party frameworks rather than only Google's own stack, which is more open than vendor documentation usually manages. Now the problems, and the first is that none of this is a course.
There is no sequence, no assessment, no sense of what to learn first or how to tell when you have understood something. You are handed a reference library and left to construct your own curriculum, which is fine if you already know the shape of the field and unhelpful if you do not. Somebody arriving here with the aim of learning to build AI applications will bounce between pages, accumulate fragments, and struggle to assemble a working mental model. Compare that with a properly structured course and the difference is stark.
The second problem is churn, and it is Google's own doing. The AI product surfaces have been renamed, merged, deprecated and relaunched repeatedly, and the documentation carries the scars. You will find pages describing an approach that has since been superseded, guidance that points at a product with a different name than the one in your console, and enough overlapping entry points that working out which documentation applies to your situation is itself a task. This is not incompetence so much as the cost of moving fast, but it lands on the reader.
The third is the usual vendor limitation. Everything is framed as though Gemini is the obvious answer. There is no honest discussion of where another provider's model would serve you better, no comparison of cost or capability against the alternatives, and no acknowledgement that most serious applications end up provider agnostic for good reasons. You need to get that perspective elsewhere, and you should, because architecting yourself into a single provider on the strength of that provider's own documentation is a decision you will regret at renewal.
And it assumes you can already program. This is developer documentation, written for developers, and it offers nothing to somebody who has not written code before. That is a legitimate scope decision, but Google's broader marketing does not always make the distinction clear, and people do arrive expecting an on ramp. My 3.6 is for excellent examples and a genuinely low barrier to entry, undermined by the absence of any teaching structure and by documentation that struggles to keep pace with the product underneath it.
The practical recommendation is to get your conceptual foundation from a real course, whether that is a DeepLearning.AI short course, Hugging Face's material or a book like AI Engineering, and then come here for the cookbook and the reference. Used that way it is a good resource. Used as a starting point it will leave you with fragments and no framework.