The Experiment Was Simple
We generated 20 card-news captions in two different ways.
• **Method A**: a long prompt with ten lines of rules
• **Method B**: two lines of rules plus one well-written example caption
Topics, model, and character limits were identical.
The Result
With Method A, I hand-edited 11 of the 20 captions. Most had a stiff tone or ran over the character limit.
With Method B, I edited only 3.
Ten lines of rules, and it still lost. We saw similar results in a Claude Code session on a Windows environment and in a bot running on the Mac mini.
Why It Works
Rules leave room for interpretation. Even if you write 「keep it light」, how light is light differs by person and by model. An example leaves almost nothing to interpret, because it shows the standard directly.
Three Steps Anyone Can Follow
1. Pick the one result you have liked best so far.
2. Paste it as-is at the very end of your prompt.
3. Keep only the two rules you truly need and delete the rest.
Rules
• Two sentences max, under 80 characters
• Friendly tone, one emoji at most
Good example
Tip of the day: keep instructions short, attach just one example.
Write a caption for the topic below in the same tone as the example above.
Things to Watch
• Start with **one** example. Past three, the output starts copying the example sentences.
• Choose a topic different from the example. With the same topic, even the content starts to look alike.
• For multilingual work, keep a separate example per language. Forcing one Korean example to cover English and Chinese makes the result feel unnatural.
Next time you feel the urge to add another rule, attach one example first. It finishes much faster than fixing ten lines.