---
id: examples-and-constraints
title: Examples and limits in prompts
level: 1
modes: [day-pack, rucksack]
moment: "An ask came back generic and one example and one limit would fix it, or they open an ask with a persona such as 'you are the world's best marketer'."
keywords: [example, one example, limit, word limit, constraint, tone, format, copy my example, show don't tell, few-shot]
source: "research/2026-10-20-concepts/examples-and-constraints.md; research/2026-10-23-essentials/batch-2-facts.md; mapping (product/learning-system/MAPPING-1-2.md): FAQ 1.09; MISTAKES 1.10 (home), 2.4; tips T023, T024, T026; strand A"
---

## What it is
One example of the result you want, such as a real message you sent before, often steers an AI app better than adjectives like "make it punchy" or a persona like "you are the world's best marketer". One limit with a number, such as "under 80 words", steers it the same way. Say what to borrow from the example, such as its tone and length, or your AI may copy its wording and facts.

## Tips
- Give your AI app one example of what you want and say what to borrow from it, because it may take the example's wording and facts as the pattern.
- Add one limit with a number, like under 80 words, because a limit closes a gap your AI app would otherwise fill by guessing.
- Give your AI app the task's facts and one example instead of "you are an expert", because a persona can change tone but cannot supply facts it was never given.

## Questions people ask
- Q: How many examples should I give? A: Start with one example and say what to borrow from it. When the format matters, give a few varied examples, which teach the pattern without your AI copying any one of them.
- Q: Should I tell it it's an expert? A: It may change the tone, but studies found an expert persona did not make answers more accurate. Give it your facts and one example instead.
- Q: Why did it copy my example? A: Your AI reads the example's wording and facts as part of the pattern. Say what to borrow, such as same tone and length but new facts, or use an example with the details changed.
- Q: Why did it ignore my word limit? A: AI apps mostly keep short limits but often miss on longer pieces, so give a number and trim any extra yourself.
- Q: Why does it still do what I told it not to? A: Say what you want instead of what to avoid. "Keep it under 80 words" works better than "don't make it too long".

## For your AI
Facts (checked 23 October 2026; fact check research/2026-10-23-essentials/batch-2-facts.md; research
note research/2026-10-20-concepts/examples-and-constraints.md).
- Examples: Anthropic, 10 November 2025, start with one example and add more only if the result still
  misses; say what to do, not what not to do (https://claude.com/blog/best-practices-for-prompt-engineering).
  For a format that matters, Anthropic's and OpenAI's developer guides advise several varied examples
  (Anthropic: three to five), and Google warns too many can make the model overfit
  (https://ai.google.dev/gemini-api/docs/prompting-strategies).
- Word limits: models mostly keep short limits and slip on long ones (LIFEBench, 22 May 2025,
  https://arxiv.org/abs/2505.16234).
- Personas: an expert persona had no significant effect on accuracy, and a low-knowledge one often cut
  it, on graduate-level questions (Wharton, 5 December 2025, https://arxiv.org/abs/2512.05858); 162 roles
  did not improve factual answers (Zheng et al., 2024, https://arxiv.org/abs/2311.10054). A persona can
  still set the tone. Offering a tip or making threats "generally has no significant effect" (Wharton, 1
  August 2025, https://arxiv.org/abs/2508.00614). "Think step by step" gives small gains on non-reasoning
  models and marginal ones on reasoning models (Wharton, 8 June 2025); in some apps a phrase like "think
  hard" switches on a slower reasoning mode, so never say such phrases do nothing.

Teach it like this: work from their real ask. A generic result usually means the ask lacked the task's
facts (who it is for, what is new, what they want the reader to do), so ask for those first. Ask them for
one example from the real task (last month's reply); with none, a piece they liked works, with what to
borrow named and their own facts in place of its facts. Add one limit with a number that suits the job. If
they reach for a persona, say it may set the tone but will not make the facts more accurate. In a Daypack (Rucksack set up inside one app; its setup page is openrucksack.com/start.md),
the plan's example tip is "When a tenant reply comes back generic: paste last month's as
the example" (section 5).