The Essential Guide to Prompting

핵심 요약

• Six ideas explain almost all of prompting: intent, context, structure, examples, decomposition and evaluation, and each one fixes a specific kind of failure.
• Context is a trade-off, not a virtue: too little makes the tool guess, too much buries the task, so give only what changes the answer.
• Examples are the strongest lever you have and also the stickiest, because the tool copies their flaws as faithfully as their strengths.
• Break any job with more than one judgement into steps, and let the tool critique its own draft against a rubric before you read it.
• Prioritise a clear task, a tested rubric and a few good examples; ignore magic phrases, elaborate personas and anything you cannot measure.

이 비트의 내용

1. The map: six ideas that cover nearly everything

   Prompting advice arrives as hundreds of tips, most of them true and none of them organised. Underneath them sit six ideas.

2. Intent and context: the input side

   Intent sounds obvious and is the idea most often skipped. 'Summarise this' is not an intent; 'give me enough to decide whether to attend this meeting' is.

3. Structure and examples: the output side

   Structure is the shape you ask for: a table with named columns, a numbered list of exactly five, a paragraph then a one-line recommendation, a JSON object with fixed keys. Naming the structure costs one sentence and removes most of the editing afterwards.

4. Decomposition: one judgement per step

   A prompt that asks the tool to research, decide, write and format in one breath asks for four judgements at once and gets each done at a quarter of its ability. Decomposition splits the job so that each step needs one judgement.

5. Evaluation: how you actually get better

   Without evaluation, prompting is superstition: you change a word, the answer improves, and you cannot tell whether the word or the dice did it. Evaluation starts with a rubric, three to five statements a good output must satisfy, written before you look at the reply.

6. What to prioritise and what to ignore

   With the map in hand, priorities are clear. Spend your effort on a named intent, on context that would change the answer, on one deliberately chosen example, and on a rubric you test against.