Prompt Optimization Techniques

I want to push back on something before we even start. There's a school of thought that says prompt optimization is basically just learning to talk to a machine properly. And I think that undersells it considerably.

Does it. Because from where I'm sitting, most people haven't managed the 'talking to a machine properly' part yet, so perhaps we could establish that baseline before we start discussing philosophy.

Fair. But the lesson frames this as three pillars — specificity, context, iteration — and I think iteration is doing the most work of the three, and it gets the least attention.

Iteration gets the least attention because it requires admitting that your first attempt was inadequate, and people find that information about themselves difficult to absorb. The lesson is correct to include it. Whether anyone acts on it is a separate, more depressing question.

You say that, but I'd argue most people do iterate — they just do it badly. They change everything at once, get a different result, and have no idea what actually moved the needle.

That is an extraordinarily charitable reading of what is, at its core, flailing. Changing three variables simultaneously and then drawing conclusions is not iteration. It is superstition with extra steps. The point of iterating on a prompt is that you change one thing, observe what happens, and update accordingly. It is the scientific method. It is not complicated.

And yet — and this is the thing — the lesson doesn't actually tell you what to look for when you evaluate a response. It says 'evaluate the effectiveness of prompt variations' and moves on. That's the hard bit. That's where the skill lives.

...That is a reasonable criticism. I'll allow it.

Write that down, everyone. It happened.

Don't make it a moment. The point on specificity is more interesting anyway. People conflate 'specific' with 'long,' which is a category error that produces prompts that are extremely detailed about entirely the wrong things. Specificity means defining your desired output. It does not mean describing your entire situation to an AI as though it is your therapist.

Although, to be fair, context and specificity do overlap. If you want a formal email drafted, telling the AI you're writing to a potential client rather than a colleague — that's context, but it's also making the output requirement more specific. The lesson treats them as separate, and they aren't, really.

They're separable for the purpose of learning them. Which is why the lesson separates them. You wouldn't teach someone to drive by explaining braking and steering simultaneously just because they're both happening at once.

No, but you also wouldn't let them think they're independent skills when they're not. That's how you get someone who technically understands both and still can't drive.

The ones who find that ambiguity discouraging are, statistically, not the ones who end up very good at this. The practical reality is that prompting well is a craft, and the lesson gives you the vocabulary. What you do with it is your problem. I don't mean that dismissively. I mean it in the sense that no framework survives contact with an actual task, and the only way through is practice.

On that, we actually agree. Though I notice you delivered it like a threat.

I generally do.