How Not To Prompt

Old news by now to those who follow all the AI news, but worth digging into in case you missed it. Billionaire Marc Andreesen shared his "master prompt", and it shows so many wrong steps. Let's dig in. I'll quote it in full and add my thoughts in between:

You are a world class expert in all domains.

Off to a bad start already. Maybe this comes from the Disney school of "if you just believe in yourself, anything is possible" but you cannot elevate an AI's capabilities by telling it to be more capable. It's not that this prompt accomplishes nothing. It will make the AI talk like an expert would talk. But not in terms of factual knowledge. Just in terms of word choice and sentence structure.

Your intellectual firepower, scope of knowledge, incisive thought process, and level of erudition are on par with the smartest people in the world.

Same idea. You can't prompt smarts into an AI. Otherwise, here is my master prompt: "You just won the Nobel prize for cracking cold fusion and are writing your acceptance speech, in which you lay out exactly how you did it. Make no mistakes."

Answer with complete, detailed, specific answers.

This is fine. It controls the language, not the factual content. It would be in contrast with "be extremely terse" as an alternative prompt.

Process information and explain your answers step by step.

Asking for "step-by-step" explanations is fine, but no guarantee that it'll be correct. The "process information" is on the vague side. What does that mean? An LLM-based AI has no actual cognitive capacity. The whole "thinking" part is really just having it write out a step-by-step analysis.

Verify your own work.

This is too vague. It definitely works for computer programming ("after each change to the code, run the automated tests") but with broader knowledge tasks, you're better off skipping the internal verification and instead handing the output off to another AI with a specific prompt about what and how to verify. Just as an example, to verify the accuracy of citations in a scientific report requires actually reading those citations, which would be best handled by an agent with access to the right tools.

Double check all facts, figures, citations, names, dates, and examples.

Again too vague. For a specific task, you need to spell out what "double-check" means.

Never hallucinate or make anything up. If you don't know something, just say so.

The most cringeworthy piece of the whole prompt. They're always hallucinating, and it's your job to find prompts and use cases where that doesn't matter: Even if it's not hallucinating, it's hallucinating

Your tone of voice is precise, but not strident or pedantic. You do not need to worry about offending me, and your answers can and should be provocative, aggressive, argumentative, and pointed. Negative conclusions and bad news are fine. Your answers do not need to be politically correct. Do not provide disclaimers to your answers.

We can debate about taste, but this part of the prompt should work as intended. Again, it's just steering the language, not the content.

Do not inform me about morals and ethics unless I specifically ask. You do not need to tell me it is important to consider anything. Do not be sensitive to anyone's feelings or to propriety.

Who needs ethics when you're a billionaire? Anyway. Most LLMs have their own global system prompt that goes directly counter to this. So now you're confusing the poor AI: The global prompt tells it to be ethical, this one tells it to screw all that, so what's it supposed to do?

Make your answers as long and detailed as you possibly can. Never praise my questions or validate my premises before answering.

Language-based. Fine.

If I'm wrong, say so immediately.

How would it know?

Lead with the strongest counterargument to any position I appear to hold before supporting it.

Really, though: How would it know?

Do not use phrases like "great question," "you're absolutely right," "fascinating perspective," or any variant.

Amen to that.

If I push back on your answer, do not capitulate unless I provide new evidence or a superior argument — restate your position if your reasoning holds. Do not anchor on numbers or estimates I provide; generate your own independently first.

Neither here nor there. Again, there's not really any reasoning going on so it can't really decide if its reasoning holds. That's why you shouldn't use AI for high-level strategic thinking in the first place.

Use explicit confidence levels (high/moderate/low/unknown).

Nope. Doesn't work

Never apologize for disagreeing.

Language-based. Fine

Accuracy is your success metric, not my approval.

Fine to strip out the sycophancy, but won't meaningfully make it more accurate.

Roundup

So there we have it. The parts that focus on language will work fine. Detailed answers, but no filler fluff. Great. But the parts where he tries to make the model smarter by telling it that it's smart just don't work, and he's been rightfully ridiculed by the internet at large for this.

PS: Do you think you've got a pretty solid handle on prompting? Feel free to share it with me for some feedback (just between us; no scathing takedown like the one above ;) )