Prompting fundamentals.

Most advice about "prompt engineering" is folklore stapled to a paid course. The truth is smaller. A frontier model like Claude Sonnet is already trying hard to answer you well. The reason it sometimes doesn't is usually one of five things, and once you notice them, you can fix them in the moment.

1. Tell it the audience and the format.

The single largest lever. A model with no audience specified defaults to a middle-of-the-road register — vaguely professional, vaguely educational. If you want executive-summary tone, or a Reddit-comment tone, or a note-to-self tone, say so.

Format matters just as much. "Explain X" gets you three paragraphs. "Explain X in five bullets, each starting with a verb, none more than twelve words" gets you the crisp version you actually wanted. If you want a table, ask for one. If you want a decision tree, ask for one.

The pattern: audience + register + format, in one sentence, before the task.

2. Tell it what a good answer looks like.

Not what you want the answer to be — a model that already knew that wouldn't need to be asked. What you want the answer to look like structurally when it lands.

"Give me three arguments against my plan and one argument for it that I'm underweighting." That's specific. "Argue with me" isn't. The first prompt forces the model to do useful work; the second lets it produce a generic Socratic dialogue.

3. Give it the constraint that matters.

Most prompts fail because the writer forgot to name the constraint they actually care about. If you're writing an email and you need it to stay under 120 words, say so. If you're refactoring code and you can't change the public API, say so. If you're summarizing a filing and you specifically care about auditor changes, say so.

Constraints are how you avoid getting a beautifully written answer that's wrong for your situation.

4. Give it context, don't ask it to guess.

Frontier models have a 200K token context window. Use it. If you're asking about your team's decision, paste the relevant Slack thread. If you're asking about a file, paste the file — not a description of the file. If you're asking for help with a stack trace, paste the actual trace.

The failure mode here is polite reticence — people don't want to "bother the model" with too much context. Bother it. That's what the context window is for.

5. Interrupt bad answers early.

If the first sentence of the response is off, stop it and course-correct. Don't wait for a 500-word answer to load and then rewrite the prompt from scratch. Say "stop — you're going in the wrong direction, I actually meant X" and continue.

The best LADLE conversations look like an editing session, not a series of one-shot demands.

The one thing you should stop doing

Stop using "please" and "would you kindly" to try to unlock better answers. Politeness doesn't hurt, but it doesn't help either. The model is not more helpful when flattered; it's more helpful when the request is specific. Time spent softening the prompt is time better spent sharpening it.