常见的提示词问题及其规避方法

Even a capable LLM fails on a fuzzy prompt. Most disappointing results trace back to a few avoidable mistakes - here they are, each with a fix.

"Write something about marketing" gives the model no anchor. Always specify the topic, audience, length, and format, as the good example shows.

Cramming five unrelated asks into one prompt yields shallow answers to each. Break the work into focused steps or separate prompts.

If you need structured output, say so. Without guidance the model invents its own format, which is painful to parse - spell out the schema, like below.

Telling the model only what not to do leaves a vacuum. Pair each prohibition with a positive instruction, as the better example shows.

Words like "it" or "that" confuse the model when context is long. Restate the actual subject explicitly so there's no ambiguity.

When you want a specific style, show an example instead of describing it. One demonstration usually beats a paragraph of rules, as below.

Dumping huge irrelevant text dilutes attention and burns tokens. Include only what the task needs, and put the key instruction near the end.

LLMs state wrong facts confidently. For anything factual, ask for reasoning or sources - and verify critical claims yourself.

The first prompt is rarely the best. Treat prompting as a loop: read the output, spot what's off, and refine one variable at a time.

A quick checklist: Is the task specific? Did I state audience, length, and format? Did I show an example? Is context trimmed? Will I verify the facts?

Recap: weak results come from vagueness, multitasking prompts, missing formats, negative-only instructions, and not iterating. Be specific, show examples, trim, verify, refine.