Lab Notes · 5

Short Prompts vs. Long Prompts — The Break-Even Point on Length

Jun 16, 2026 · AI Note Lab

Prompt length vs. quality — the curve peaks around 300 characters
Prompt length vs. quality — the curve peaks around 300 characters

"The more detailed the prompt, the better" — I've said it on this blog more than once. But how detailed is detailed enough? Is longer always better, without limit? I ran the same task through five prompts, changing only the length.

Setup

Fixed task: "write Instagram promo copy for a neighborhood bakery's new item (a salt bread roll)." I prepared prompts at five length tiers.

  1. ~10 characters: "Write promo copy for salt bread"
  2. ~50 characters: + it's a neighborhood bakery, and it's for Instagram
  3. ~150 characters: + target audience (20s–30s), tone (friendly), length (2 sentences + hashtags)
  4. ~300 characters: + the shop's personality (dough mixed at 4:30 a.m., French butter), phrases to avoid (no "crispy outside, soft inside")
  5. ~500 characters: + 2 examples of good copy, the brand story, event details, an expanded list of banned phrases

Results

TierVerdict
10 charsCopy you could paste onto any bakery. In other words, not ours
50 charsStill generic. Mostly stock phrases like "freshly baked" and "a happy bite"
150 charsThe big jump happens here. With length and tone pinned down, it's postable as-is
300 charsThe shop's unique details (pre-dawn dough, the butter) become the heart of the copy. Peak quality
500 charsAbout the same as 300, or subtly stiffer. It felt boxed in by the examples' style

Observation: a curve of diminishing returns

Quality did not scale with length. It climbed steeply from 10 to 150 characters, peaked at 300, then flattened — or dipped slightly. Digging into why the 500-character tier dipped, I found two reasons.

What I learned

The practical rule of thumb: fill in roughly three sentences — what, for whom, and what's uniquely ours — and your prompt becomes "long enough" on its own. Don't count characters; count pieces of information.
← Previous experiment · Image AI and abstract instructions