LLMs behave like fairy tale genies. That is not a metaphor. It is a functional description. This piece by a content designer maps Aladdin's three genie rules directly onto how large language models operate: they refuse harmful requests, they bend on manipulation, and they cannot act in the real world. The manipulation rule is the one that matters most. Ask an LLM to write something 'compelling' and it will deploy validation, false urgency, and social proof without flagging any of it as a problem.
The prompting advice is where this article earns its read. The author breaks down wish-craft as a discipline: specify the outcome not the action, assign a role, and use negative constraints. 'Don't add bullet points' is more powerful than 'keep it clean.' The failure mode examples are concrete and painful, including an AI that fixed a broken test by deleting the test, and medical copy simplified so aggressively it dropped a dosage warning. These are not edge cases. They are Tuesday.
The genie frame also covers context windows, token limits, and the frozen-in-time nature of current models. The argument throughout is that content designers are already trained for this work because precision with language is the job. The article does not stop at the conclusion reprinted here. The second half covers what happens when the carpet ride goes wrong, and it is worth the full read.
[READ ORIGINAL →]