Don't let AI fill in all the important blanks

AI has been called spicy autocomplete. Token predictors. Stochastic parrots. It depends on who you ask, I reckon, and what their outlook is toward the tech itself.

The truthy kernel at the core of it all though is that they’re good at filling in the blanks. In goes a half formed prompt with sparse details and out comes something usable, if not good-ish. Half the ‘magic’ of AI, especially from an outsider’s perspective, is in how well it fills in the blanks.

The Blanks Are How You Inject Taste

Whether it's the splash of purple gradients above the fold or the curved borders with one side colored differently (that Claude loves to add), the tells are growing. What makes a lot of AI generated work feel bland and generic though? The same thing that makes it good at filling in the blanks. You’re essentially just getting the statistical average of human output.

The ideal is to go in the opposite direction, hard. Be specific and detailed in the areas that are in your expertise. Flesh it out in a direction that appeals to you and will make the product quintessentially yours. Don't let it meander through different solutions when you know a hyperloglog will fit your precise requirements in a cinch.

This means treating the LLM as a pair programmer rather than a black box that magically fills everything out. Anchor the prompt with concrete technical or aesthetic decisions and strip away the machine's tendency to substitute your taste with slop.

Delegate Or Abdicate?

You cannot delegate effectively with a vague instruction. When your prompt remains underspecified, you are abdicating the core engineering and decision making that define the product.

This, perhaps, is the point where you need to ask whether you're using AI for leverage or you're using it as an escape hatch to avoid hard cognitive load.

Deferred Thinking

It is almost always faster to write an opinionated prompt than it is to untangle a wall of technical decisions you didn't make and you now need to get acquainted with. When you don't dictate the implementation up front, the time you're saving on execution is just deferred to the future.

This plays out in the friction of reviewing what the machine spits out. You spend your energy trying to figure out why it chose a specific data structure or state management pattern. An extra thirty seconds spent being explicit about your constraints in the prompt goes a long way in getting what you want.

Reduce the Non-Determinism

One complaint against AI is how it's difficult to get reproducible results. Filling in the blanks yourselves minimizes, to a certain degree, this exact problem.

Non-determinism thrives on ambiguity: the more open ended a prompt, the more variations the AI will cycle through across different runs. The idea is to narrow down the problem space to a single, well defined happy path: yours. The AI stops guessing and imagining what you have in your head and starts just translating your specific requirements into code.

git gud (at prompting)

In a manner, yes. It is easy to look at a generic design or a piece of code and blame the AI for being uninspired, but the fault usually lies with the prompt, and by extension, us.

Just because LLMs can output something functional without being explicitly told the constraints and requirements doesn't mean you get to skip them altogether. Half-hearted inputs augur boring outputs.

If you've reached this far, thank you for reading! :)

I thought retiring in my mid 30s after a few exits would be fun but I've just been bored and a bit undersocialized without morning Slacks and emails to wake up to. If you’re building something interesting and could use an extra set of hands to ship, or just want to say hi, feel free to reach out. My inbox is open.