Calibrating tone with an example means attaching an old email whose tone was pitched exactly right as a few-shot example, letting Claude imitate the specific features that make up tone — word choice, sentence length, forms of address — rather than trying to define an abstract requirement like 'professional but not too stiff' into a set of written rules. This differs from the usual instinct in prompt writing: most people's first move when facing a tone requirement is to write it out in more detail, but tone is a holistic impression made of many details that resist being enumerated, and writing it as rules tends to get longer and vaguer without ever guaranteeing the combined result actually feels right. An example bypasses the 'define it' step entirely, converting a standard that's hard to put into words into something concrete Claude can directly imitate.
This is needed because when a manager hands off a tone requirement, it usually comes as a short phrase like 'professional but not too stiff' — a phrase that's actually precise for the manager, who has a clear picture in mind, but that picture can't be fully converted into words and passed to you, and when you pass it on to Claude unchanged, you've just relayed a vague instruction vaguely a second time, losing information at each translation. An example cuts that chain of loss short: rather than letting an abstract requirement pass through two layers of translation — person to person, then person to AI — you find a concrete case instead and let Claude read the tone directly from it, skipping every step in between where something could go wrong.
In practice this runs in three steps. First, find an old email whose tone you think landed exactly right, favoring one closest in context to what you're replying to now — same kind of apology, same kind of proposal — the closer the context, the more comparable the example. Second, paste that letter's full text into the prompt as the example, along with the key points for this reply, changing the instruction to 'match the tone and style of this letter, and reply to the following' — no need to additionally describe tonal features, the example has already demonstrated them. Third, if a perfect match isn't available on short notice, an 80-percent match is still better than no example, since an imperfect example still carries more tonal information than a purely verbal description; if there's still a gap in the output afterward, point directly at it — 'match the tone of this particular passage in the example more closely' — a correction that's usually faster and more precise than redefining tone from scratch in words.
For you, what this approach actually saves isn't prompt-writing time — it's the time and communication cost of fixing things afterward. Get the tone wrong, and you either rewrite it yourself or take it back to your manager, who says 'that's not quite what I meant,' leaving you to guess at what's off yourself, because even the manager may not be able to articulate it precisely either — the whole back-and-forth typically costs far more than the one minute spent upfront finding a fitting old email. What's worth watching is the quality of the example chosen: if the example's context is too different from what you're replying to now — say, using a thank-you note as the example for writing an apology — Claude may pick up content features from the example it shouldn't be imitating. When that happens, the problem isn't that few-shot prompting as a method has failed, it's that the example wasn't a close enough match; swapping in one with a closer context usually fixes it.
Your manager asks you to reply to a client email and adds, offhand, 'keep it professional but not too stiff — we're on pretty good terms with this client.' You paste that instruction verbatim into Claude's prompt and get back a letter that's textbook-correct in tone and reads nothing like how you two actually talk to this client. The problem isn't that Claude can't write a letter. It's that 'professional but not too stiff' was never precisely defined to begin with — it lives precisely in the shared understanding between you and your manager, but it can't be translated into a set of generic, portable text rules.
Try writing 'not too stiff' out as a rule, and you'll find it gets longer and vaguer the more you write: don't open with a formal salutation, allow one or two colloquial phrases but not too many, sentences shouldn't run too long, but also not too short or it seems curt. Each rule sounds reasonable on its own, but stacked together there's no guarantee the resulting letter actually 'feels right' — this is a common failure mode of constraint stacking: pile on enough individual conditions and each one holds true on its own, yet the overall effect doesn't add up to their sum, and tone is exactly the kind of thing that distorts easiest under that kind of stacking. This isn't a failure of your ability to articulate things — tone is inherently a holistic impression, made up of word choice, sentence length, how you address someone, even punctuation habits, and it resists being broken down into discrete rules even though you can judge 'this one's right, that one's wrong' at a glance. This kind of standard — recognizable but hard to define — is exactly where few-shot prompting earns its keep.
The approach is simple: find an old email whose tone you think landed exactly right — one you wrote yourself, or one a colleague wrote to this same client that struck the right note — and paste that letter into the prompt as an example, along with the key points for what you need to reply now, changing the instruction to 'match the tone of this letter, and reply to the following.' You don't have to say 'no formal salutation, medium sentence length, a bit of warmth without losing propriety' — the example letter has already demonstrated all of that to Claude. The point isn't that the example needs to be perfectly written; it's that it converts the standard you can't quite articulate into something concrete Claude can directly imitate.
How you pick the example directly affects the outcome. Check whether it genuinely represents the tone you want, or whether it's just what happened to be lying around while the content isn't actually a good match. If you have several candidates, favor the one closest in context to what you're replying to now — same kind of apology, same kind of proposal, same kind of status update — the closer the context, the more the tone actually transfers. If you can't find a perfect example on short notice, an 80-percent match is still better than no example at all — even an imperfect example carries more tonal information than a purely verbal description — the same logic that governs output format specification, except what's being specified here isn't fields or structure, it's the tone itself, with the example letter acting as the concrete template for what this output should look like.
Not every email warrants the effort of finding an example. If the recipient is an internal colleague, the tone requirements are already loose, or you already have a solid sense of what this letter should look like, a clearly written verbal instruction is usually faster than digging through old emails. Where it's worth the time is when tone has a real bar to clear but is genuinely hard to put into words — replying to an important client, a carefully calibrated apology, external communication representing the company's face to the world. In these cases, getting the tone wrong is costly enough that the extra minute spent finding a fitting old email pays for itself.
'Professional but not too stiff' feels like a clear instruction, but it's really handing Claude a standard you know precisely in your own head yet can't fully put into words, and asking it to guess. Getting it right is luck; getting it wrong means rewriting, and often re-explaining what was wrong — an explanation that's usually harder to give than just finding an example letter, because you genuinely struggle to articulate exactly what's off, you can only sense that it is. Rather than spending time explaining a standard you can't quite define after the fact, move that time to before the fact instead — one fitting old email usually gets Claude to the right calibration faster than ten adjectives ever could.