Prompting Principles and Syntax

The thing that gets me — and I find quite a lot of things get me — is that people treat prompting like a vending machine interaction. Insert words. Receive output. Shake machine if disappointed.

Which is actually how most people interact with technology, so I'm not sure that's the damning indictment you think it is.

It is precisely as damning as I think it is. A vending machine has a fixed inventory. A language model does not. You are essentially shaking a machine that could make you anything, demanding a slightly better version of the worst thing in the top row.

Fine. So the lesson's point — instruction first, then context, then examples, then format — that's the alternative to shaking the machine. It's not a rigid formula, but it is a useful default.

I would go further. It isn't just a useful default. It's the difference between asking someone to 'write something' and asking them to write a specific thing, for a specific purpose, for a specific reader, in a specific register. The latter is not being difficult. It is being coherent.

Although I'd push back slightly on the examples piece. People read 'provide examples' and immediately produce mediocre ones, which is arguably worse than none. If your example is muddled, you're just teaching the model to be muddled with more confidence.

...That is a genuinely good point and I resent that I didn't make it first.

I know.

The context section is where I see the most spectacular failures. People write four words of instruction and then wonder why the output is generic. You have given the model nothing to work with. You have handed a chef an empty bowl and complained the soup lacks depth.

Though there's a counterargument that over-specifying context can constrain the model in ways you didn't intend. Sometimes the brief being slightly open is what gets you something interesting.

That is true for creative tasks and almost nowhere else. If you are asking an AI to draft a client-facing document or extract structured data, 'slightly open' is not a creative strategy. It is an abdication.

Agreed. The lesson doesn't really distinguish between those two modes, which I think is its one weak spot. The principles are solid, but they're presented as universal when really the weighting shifts depending on what you're trying to do.

A fair criticism. Though I'd argue that understanding the principles well enough to weight them appropriately is exactly the skill the lesson is trying to build. You cannot calibrate a framework you haven't internalised.

So the actual takeaway is: learn the structure, then learn when to break it. Which is, annoyingly, how most useful skills work.

Annoyingly, yes. People who find that discouraging are not, in my experience, the ones who become particularly good at this.