**Few-shot example order** matters because LLMs read left to right and the last example sits right next to where the model starts generating. The correct answer is B.
This is recency bias. Attention is global but practical attention weight skews toward nearby tokens, and the tokens immediately before the model's first generation step are the ones with the strongest pull on the output distribution.
The practical rule has two parts. First, put your most-representative example last. If you have five examples and one is the canonical case (typical input, ideal output, the pattern you want the model to imitate), that one belongs in the last slot, not the first. Second, avoid style oscillation. If you alternate verbose, concise, verbose, concise, the model often averages or flips between styles. Keep the style direction consistent across examples and let the last example anchor it.
The other options are common traps. Option A repeats the folklore that 'examples are a set'; multiple papers measured the effect size at 5 to 15 points on benchmarks, which is huge. Option C orders by length; length is not the right axis. Option D orders alphabetically; that is arbitrary and does not connect to where the model generates from.
**Few-shot order** matters because of recency bias: the model generates right after the last example, and that example primes the next-token distribution most strongly. Answer B.
The rule is to put the most-representative example last so the cleanest pattern is the one immediately adjacent to the model's generation cursor. The second rule is to avoid style oscillation; alternating contradictory examples (long, short, long, short, or formal, casual, formal, casual) causes the model to average or flip between modes.
Measured order effects on standard benchmarks land in the 5 to 15 point range, which is enough to flip a regression test. That makes example order a real hyperparameter, not an aesthetic choice.
The other options miss the mechanism. Order matters (A is wrong). Length is not the axis (C is wrong). Alphabetical order is arbitrary relative to the position that matters, which is 'distance from the query' (D is wrong). The only ordering that connects to how the model uses examples is by representativeness, last slot first.
Quick reference
TL;DR
LLMs are recency-biased; the example nearest the query has the strongest priming effect, so put the most-representative example last and avoid style oscillation across examples.
Imagine you are about to imitate someone's drawing style. Someone shows you five sketches, then says: 'now you draw.' The sketch you saw most recently is the one most fresh in your mind, so it will shape your strokes the most. If they show you a cartoon, then a realistic portrait, then a cartoon, you will probably mash styles together. Few-shot examples work the same way: the model is about to draw right after the last example, so put the cleanest one closest to the prompt, and do not zig-zag between styles.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Tasks where the model is already at ceiling without examples Few-shot adds tokens and ordering risk. If a zero-shot prompt already passes the golden set, stop there.
Open-ended generation where you do not want the model to imitate a specific style Examples bias the model toward their pattern. For creative or exploratory tasks the priming becomes a constraint, not a help.
Massive example libraries that change frequently Hand-ordering does not scale; use retrieval-based selection (DSPy, similarity ranking) and let the system order by similarity instead.
Quick scan · 2–4 min
Everything important, quickly.
3 min: recency bias mechanism + canonical example last rule + style oscillation failure + measured effect size + interaction with long context and retrieval.
Real-world examples
Real products, models, and research that use this idea.
1Anthropic's prompt engineering guide explicitly recommends placing the strongest example last for Claude Opus 4.7 and Sonnet 4.x few-shot prompts.
2DSPy automatically reorders selected demonstrations by similarity to the query, with most-similar last, when compiling prompts.
3OpenAI cookbook examples for few-shot classification show the canonical (input, output) pair as the final demonstration for the same reason.
4LangChain's `FewShotPromptTemplate` exposes an `example_selector` slot that ranks examples; teams pair it with similarity-based selection putting the closest match last.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QDoes the recency bias rule still apply when the prompt is 32k tokens long and examples span the middle?
Sign in or go Pro to see the approach for this follow-up.
QIf you have ten examples but the model context can only fit four, how do you pick and order them?
Sign in or go Pro to see the approach for this follow-up.
QDoes example order matter as much for instruction tuned models like Claude Opus 4.7 as it did for base models?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Treating the example list as a set rather than a sequence. The model reads it left to right and the last example is right next to where it starts generating; that position matters.