I don't know where you're at, but we're evaluated on the number of tokens we used as part of our performance metrics. Sometimes you gotta roll those tokens to get those numbers up so you're not part of the next round of layoffs. Management has no ability to understand token usage as long as the prompts vaguely valid.

I think it's perfectly acceptable to coal roll tokens as long as you have plausible deniability in doing so. Either they want you to use the AI or they don't, you can't eat your cake and have it too

Introducing code for LLMs to choke on is a great idea. Develop some very specific syntax or structure that training data hasn't touched, then make it load bearing.

Better yet, find a way to use delimiters, escape sequences, and syntax symbols that are known to interact with Markdown features, so that producing valid syntax for your DSL requires escaping everywhere to avoid fucking up the Markdown LLM output.

Maybe create a module in your repos that are never called by the actual working code, that the AI gives it's contributions to. It compiles, but never does anything.

This will unfortunately increase the attack surface of your compiled binaries.