Writing Clear Instructions

Why Specificity Matters

A language model tries to produce the most probable continuation of your prompt. A vague prompt like "write something about climate" has an almost unlimited solution space — the model can produce a poem, a scientific essay, a news article, a listicle, or a children's explanation, all of which are valid interpretations. The more you constrain the solution space with specific instructions, the more reliably the model lands where you actually want it.

The Four Dimensions of Clear Instructions

Great instructions specify four things: what (the task), who (the intended audience), how (format and structure), and how much (length, depth, number of items). "Write a 200-word summary of the three main causes of climate change, suitable for a 12-year-old" hits all four. Compare that to "explain climate change" — only the "what" is present, so the model has to guess the other three and will often guess wrong.

Format Directives

Telling the model how to format its response is one of the highest-leverage changes you can make. "Return a bullet list" vs "return a numbered list" vs "return a table with columns X, Y, Z" each produce completely different structures. If you need the output to feed into code, say "return a JSON object." If you need it to paste into a document, say "use markdown headers and bold key terms." The model follows formatting instructions with high reliability when they are stated explicitly.

Common Mistakes to Avoid

• Assuming shared context: The model doesn't know what "the document" is unless you paste it. Don't refer to things that aren't in the prompt.
• Over-hedging: Phrases like "if possible, maybe try to sort of…" dilute the instruction. Be direct.
• Contradictory constraints: "Be brief but cover everything" gives the model an impossible trade-off. Pick one.
• Missing the negative space: If there are things you explicitly don't want, say so. "Do not include statistics" or "do not recommend paid tools" are instructions the model will follow if given.