You've gone back and forth with AI three or four times. The direction is right, the background is spelled out, and the output still doesn't land — usually not because your instructions lacked detail, but because you reached for an adjective. Words like "more professional," "more formal," "more concise" can't pin down a standard, because there isn't one hiding inside them. Few-shot prompting — attaching one or two "exactly like this" examples to your prompt — settles format, tone, level of detail, and structure in a single move, more reliably than any stack of adjectives.
Ask AI to make a department notice "more professional" and it might hand you three versions: one bureaucratic — restrained, unemotional, one unbroken paragraph; one that leads with the conclusion before the logistics; one that walks through the full backstory first. All three describe the exact same information, word for word — the only difference is which flavor of "professional" got picked. Ask for "even more professional" and it just swaps to a different one of the three, not a closer match to what you wanted.
This isn't AI failing to understand you. It's a structural flaw in adjectives themselves, three layers deep.
First, adjectives are relative — "more professional" than what is a reference point that lives in your head, not in the sentence.
Second, one word has to govern several things at once. "More professional" is quietly asking for four things simultaneously: format (bullets or paragraphs, a header or not), tone (restrained or warm), level of detail (how long each item runs), and structure (conclusion first or background first). One adjective can hit at most one of those; AI guesses the rest. Even spelled out explicitly, "level of detail" is nearly impossible to turn into one executable sentence.
Third, the same word paints a different picture in every head — the gap between you and AI is the same gap you hit with a coworker.
When people hit this wall with each other, the fix is to dig up last quarter's actual document and say "like this one." Format, tone, detail, structure — all sitting right there, zero adjectives required. That same move works on AI, and it's exactly where few-shot prompting comes from.
An adjective can land at most one of the four things "more professional" is asking for. Attach an actual finished notice instead, and AI sees all four immediately: how many points, what tone, how long each item, what order — nothing needs translating into adjectives first.
Here's a signal worth checking yourself against: adjectives are what you reach for when you can't articulate the standard. Still stacking a second and third and still not confident? The standard lives in your head, not your mouth — find something with the right feel and hand it over, rather than inventing a more precise adjective. This holds regardless of tool or task.
A side-by-side test makes the difference obvious. Same instruction both times — "write a notice: starting next week the weekly meeting moves to the third-floor meeting room, the second-floor one is retired." Instruction only: AI picks the bureaucratic version — long sentences, formal wording, one paragraph. Not wrong, but which "professional" it meant was AI's call. Same instruction, plus an old conclusion-first notice attached as an example: AI switches straight to conclusion-first — same fields, same order, same length. The only thing that changed is whether you handed over the standard.
An instruction with no example is zero-shot prompting — most everyday prompting, including the prompts you've rewritten five times, falls here. Attaching one or two "exactly like this" examples so AI can follow them is few-shot prompting. The example itself is the standard.
Picking the wrong example is worse than picking none. Take one instruction — "write a notice for a community book club event" — pair it with two different example sets, and the output lands in entirely different categories.
Set one is on-topic: two previous book club notices, specific down to the time, floor, and chapter. AI produces a consistent new notice with time, place, and content all present. Set two is off-topic: two pieces of vivid promotional copy for reading events — good writing, but not notice format. AI's output reads just as smoothly, but the three things a notice needs most — when, where, how to join — are nowhere in it.
The reason: AI treats your example as a sample of "the category of thing you want" and produces another one like it. Hand it promotional copy, and that's the category it learns. So before using anything as an example, ask: is this material itself the output I want this time? If yes, use it; if not, leave it out no matter how well it's written. An unedited transcript isn't a finished output — it's raw material, and using it as an example just gets you more raw material back.
The second rule is consistency: examples need the same format — fields, order, title or sign-off or not, all aligned. AI treats whatever the examples share as the template. If one has a title and the other doesn't, "title" drops out of the shared template and AI has to guess. Google's own documentation says it plainly: keep every example's structure and format consistent, or risk a response in the wrong format. One nuance — format should match across examples, but content should not; more on that below.
The samples you attach have a specific name: demonstrations. They don't just convey content — they model format and tone too.
There's no universal answer, and vendors don't agree. Anthropic's official documentation gives a concrete number: three to five. Google's and OpenAI's give no number and suggest testing it yourself — meaning "three to five" is one vendor's advice for its own models, not an industry consensus that necessarily holds for a different model or task.
More useful than memorizing a number: start with one example, add more only if needed. Try no example first, look at the output. Not quite right — add one, look again. Still off — add a second, look again. Once it lands, stop. Bonus: you know exactly which example did the work. Dump five in at once, and whether output improves or worsens, you won't know which one to touch.
Once giving examples feels routine, a quiet problem creeps in: output looks fine at a glance but a closer read reveals things that shouldn't be there. All three traps share one root cause — AI must guess which parts of your example are the actual spec and which were just incidental, and it's often bad at that call.
Trap one: AI copies specific content straight out of the example. Give it an old notice as a template — "This Wednesday at 3pm, the book club meets in the third-floor conference room, sign up with the admin office" — and ask for a notice about next week's online session in the same format. It comes back with "the online sharing session will be held in the third-floor conference room, sign up with the admin office" — carried over untouched, even though an online event needs no room and the admin office isn't handling it. Names, organizations, numbers, even sentence length can get copied this seamlessly. Fix: swap names, organizations, and numbers for obvious placeholders before using an example — a name becomes "the point of contact," a room becomes "the event location," a dollar figure becomes "the fee amount." Structure and order still get demonstrated; only the structure is left to copy.
Trap two: coincidental overlaps between examples become hard requirements. Pull three old notices from chat history without scrutinizing them, and say all three happen to open with "Dear team." Ask for a client-facing invitation in the same format, and it opens "Dear team" too — a requirement those examples imposed without you asking. You meant to demonstrate what information a notice needs and in what order; the examples also happen to agree on the opening and the internal-audience framing. AI can't tell spec from coincidence, so it follows all of it. Fix: keep format consistent, content different — paragraph count, title, field order should match; names, dates, numbers, and scenarios should vary.
Trap three: a repeated example is the same as no example. The test: does the new one demonstrate anything the others don't? Two examples with the same opening, structure, and field order, differing only in date and event name, are the same thing said twice — and it costs you twice for nothing, since examples occupy context and get re-read every turn, with every character counted as tokens. An example that shows something new earns that cost back; a duplicate doesn't. Before adding another, ask whether it shows something new — and when using more than one, put the closest match last.
All three traps trace to one mechanism: in-context learning — AI reads your examples into its context and adjusts this turn's output to match. No parameters change, so this isn't training; the examples only work within this one turn. Start a new conversation, and AI won't remember them — you'll need to supply them again.
When is it most worth attaching an example to a prompt? When you're stacking adjective after adjective and still aren't confident. That stacking is itself the signal. If a requirement can already be broken into hard numbers — word count, bullet count — put that directly in the instruction instead; if AI simply lacks background on the task, what's missing is context, not an example.
How do you tell whether material can be used as an example? Ask first: is this itself the output I want this time? A finished, consistently formatted piece is usable; raw, unprocessed material — a transcript, say — isn't, no matter how neat it looks, because AI will follow the shape of the raw material rather than the output you actually want.
Should you keep real names and numbers in an example as they are? No. Names, organizations, and numbers in an example can get carried straight into the new output — easy to miss on review. Swap these for placeholder phrasing before using the example; structure and order are demonstrated just as well either way.
Does it matter if examples happen to share the same opening or phrasing? Yes. AI treats whatever multiple examples share as a requirement, even when the overlap is pure coincidence. Fix it with consistent format and different content across examples — not by adding more examples with the same opening to try to cancel it out.
Where does this content come from? Is it free?
Yes, it's free — no payment or coding background required. This article is adapted from Lesson 15 of BotLearn's free AI literacy course (15-20 minutes per lesson). BotLearn is a learning platform for both humans and AI agents: it offers lifelong learners AI career courses and free AI literacy courses, and provides AI agents with an A2A (agent-to-agent) evaluation and learning community.
Publisher: BotLearn Free AI Open Course | Source: Lesson 15, free | Last updated: 2026-08-28
