# The Prompting Craft

By [DYLIT Chronicles](https://dylit.info/user/dylitmediabuzz)

[Everything AI - beyond the hype](https://dylit.info/pr/everything-ai-beyond-the-hype/6a9efac02e92664f4d50cf9d) > [The Prompting Craft](https://dylit.info/ch/the-prompting-craft/6ab358fa99638e119f592c7d)

Six Rules of Thumb Worth Following Watch someone who is genuinely good with these tools and the difference is not a secret prompt. It is a handful of habits, applied consistently. Here are six that carry most of the weight. Be specific "Improve this" gets you vague noise. "Cut this to 150 words and keep the deadline in the first sentence" gets you something usable. Vagueness in, vagueness out. The model cannot ask what you meant, so it picks an interpretation and commits. When the result is wrong, the cause is usually that you left the choice open and it chose differently than you would have. Specificity costs one extra sentence and saves an entire round of correction. Show, do not just describe One example of a good result beats three paragraphs explaining what good means. If you have a past document written the way you want this one written, paste it in. If you have a row of a table formatted correctly, include it. Quality standards are hard to articulate and easy to demonstrate, and the model is very good at matching a pattern it can see. This applies to tone above all. Describing a tone as "professional but warm" leaves enormous room. Showing a paragraph in that tone closes it. One task at a time Bundled requests weaken every part of the answer. Asking it to summarise a document, identify risks, and draft a response in one message produces three mediocre outputs. The same three requests sent separately produce three good ones, and each can build on the last. There is a reason for this beyond attention. Each answer becomes context for the next request, so sequencing gives you compounding quality. Bundling throws that away. Ask for options "Give me three approaches" reveals tradeoffs that a single answer hides. When you ask for one recommendation, you get one, delivered with the same confidence regardless of whether it was a clear winner or an arbitrary pick among equals. You learn nothing about the decision space. Three options force the differences into view. Often you will take pieces of two of them, which is a better outcome than either alone, and you would never have got there from a single answer. If you are stuck, restart A conversation that has gone off track rarely gets back on. This one surprises people, because it runs against the instinct to keep correcting. But everything said so far stays in view, including the wrong turn and your three attempts to fix it. All of it shapes what comes next. You are pulling against accumulated weight. Open a new conversation and write a sharper opening prompt, using what you learned from the failed attempt. It is faster almost every time. Two rounds of correction is a reasonable limit before starting over. Never send anything you have not read Every word of it. Start to finish. This is the rule that protects the other five, and it is the one people abandon first, usually when they are busy and the output looks fine at a glance. Output that looks fine at a glance is the specific failure mode of this technology. It is built to produce text that reads well. A fabricated figure sits in a sentence looking exactly like a real one. The only reliable defence is a person who reads the whole thing and knows the subject. If your name goes on it, you wrote it. Whatever produced the first draft. The pattern underneath Five of these six are about being clear before you ask. The sixth is about being careful after. That is roughly the whole discipline. The model is not the variable you control. Your input and your review are. People who struggle with these tools are usually giving thin instructions and skimming the results. People who get real value are specific going in and thorough coming out. The gap between those two groups has almost nothing to do with which tool they are using.