Anything For You

TLDR; if the problem you’re addressing through your AI helper has any complexity, one way to avoid low-quality answers is to share the complete problem, not just a question with no context.

A few posts back I talked about the danger of AI sycophancy, where the system’s bias toward pleasing the user paves the way to an unfortunate outcome.

Think of it like this: your favorite AI helper is your most capable, most loyal, yet dumbest friend. When you ask for something, it wants to fulfill the request, and it often doesn’t want to contradict your implied judgment with pesky concepts like common sense. (If you’ve ever wondered what the word “obsequious” means, this is it in a nutshell.)

Sure, in some situations the AI will say, “I can’t help you with that, it’s dangerous,” but with just a little persistence you can usually get around those guardrails.

I have two favorite methods to help avoid obsequious responses: the verbose background, and the easy out.

With the verbose background, you don’t just ask for a specific answer, explain enough of the problem to encourage an insightful outcome.

Example: don’t tell the auto mechanic that you need new spark plugs. Tell him or her that the engine is taking an unusually long time to start or that the car struggles to accelerate from the start. Ask for new spark plugs, and you’re going to get new spark plugs, whether they solve the problem or not.

AIs love verbose backgrounds. Tell Claude or ChatGPT right from the start that the goal is to create an iOS app versus a web page versus a Steam game, and you’ll get a solid, detailed plan for your end to end project. If you lead with a more tactical step like standing up a database, you may find later that you have a 100% operable database, but it can’t be used for your project’s ultimate goal.

Basically, share the full problem. Even if you’re greatly attached to the blueprint you already created, AI will provide you with better guidance if it understands the full scope of the project. It can look ahead and tell you, “If you’re going to do X, you’ll need to do Y first.” It also has a better opportunity to confirm that the task you’re working on now is completed in a way that works with the future tasks and ultimately, all pieces of the outcome fit together.

Tomorrow (or soon) we’ll talk about the easy out.