{"course":{"slug":"prompting","title":"Getting Real Work out of a Model","url":"https://addaly.com/learn/prompting","level":"beginner","track":"start"},"lesson":{"slug":"capturing-a-voice","title":"Capturing a voice from three samples","url":"https://addaly.com/learn/prompting/capturing-a-voice","position":25,"of":72,"minutes":9,"answer":"Style transfers through real samples plus named mechanics (sentence length, banned openings, contractions) not through adjectives, and it drifts back to generic over long outputs.","keyIdea":"Style transfers through real samples plus named mechanics (sentence length, banned openings, contractions) not through adjectives, and it drifts back to generic over long outputs."},"body":{"format":"markdown","text":"## Adjectives do not carry style\n\n\"Professional but friendly.\" \"Warm, clear, human.\" \"Confident without being arrogant.\" Every organisation has written these words into a style guide, and none of them constrain a sentence. Two writers reading the same three adjectives produce completely different prose.\n\nStyle is one of the things prose is worst at describing and examples are best at transmitting. Three real pieces of your own writing will do more than a page of guidelines.\n\n## Choosing the three\n\n**Real, not aspirational.** The email you actually sent, the post that actually went out. Not the one you wish you had written, and definitely not something you admire from another company.\n\n**Same genre as the target.** Emails teach emails. A blog post will not teach you an email; the shapes are too different and you will get a blog post in an inbox.\n\n**Varied within the genre.** One good-news email, one difficult one, one routine one. Three cheerful examples teach cheerfulness, which is not your voice, only your voice on good days.\n\n## Strip what you do not want copied\n\nThis is the step people skip. Your three samples contain names, dates, projects, and specific facts, and some of those will come through into the new text. Replace with placeholders: `[CLIENT]`, `[DATE]`, `[PROJECT]`. Also cut any signature block, unless you want it reproduced every time.\n\nThen wrap them, so they cannot be mistaken for instructions:\n\n```\n<sample1>...</sample1>\n<sample2>...</sample2>\n<sample3>...</sample3>\n\nWrite a new email in the same voice as the samples above.\nMatch: sentence length, level of formality, how much context is\ngiven before the ask, and whether contractions are used.\nDo not reuse any facts, names or phrasing from the samples.\nSubject: [your actual task]\n```\n\nThe list of things to match is doing real work. Without it, \"same voice\" is interpreted broadly, and you get a competent pastiche. With it, the model has four concrete dimensions to attend to.\n\n## Naming the mechanics yourself\n\nYou can help enormously by describing the two or three mechanical habits that actually define your voice. Not adjectives: mechanics:\n\n- Sentences mostly under fifteen words.\n- Never opens with \"I hope this finds you well\".\n- Contractions in email, none in documents.\n- One idea per paragraph, no paragraph over four lines.\n- British spelling, Oxford comma, no em dashes.\n- Says \"we\" for the company and \"I\" for a decision.\n\nSix lines, all checkable. Combined with three samples, this is as close as you will get to reproducible voice.\n\n## The honest limits\n\n**It will drift over a long output.** Voice holds for a paragraph or two and loosens after that. For anything long, generate in sections.\n\n**It cannot invent the voice you have not written.** If you have never written a difficult message, three cheerful samples will not tell it how you would.\n\n**Generic pull is real.** Model outputs have a centre of gravity (a slightly formal, hedge-heavy, list-loving register) and long generations drift towards it. This is why the banned-words list is worth keeping alongside the samples.\n\n**Matching a voice is not matching a judgement.** The samples teach how you say things. They do not teach what you would decide to say, which is why voice-matched drafts often read correctly and land wrongly: right register, wrong emphasis, a concession you would never have made. Read for content first and style second, in that order, every time.\n\n**A voice is not authorship.** Producing text in your style does not make it yours in any sense that matters to a reader who thinks you wrote it. Where it matters (a condolence note, a reference, a personal apology) write it yourself. This is not a technical limitation and no prompt fixes it.\n\n## The free path\n\nNone of this needs a paid tool. Keep the three samples and the six mechanical rules in a plain text file. Paste them when you need them, or put them in a project or custom instruction if your tool supports it.\n\nThe file is the asset. It took you twenty minutes and it encodes something no template from the internet can, which is how you actually write.\n\n## A note for teams\n\nThe same file works for a team, and it does something a style guide never has: it settles arguments by demonstration. \"We write like this\" becomes three emails everybody has read rather than an adjective everybody interprets differently.\n\nWhere it gets uncomfortable, and where it is worth being honest, is that a shared voice file makes one person's writing the standard. Choose the samples in the open, with the people whose writing is in them, and revisit the set when the team changes. A voice imposed by a file nobody agreed to is a style guide with extra steps."},"check":{"question":"A charity wants fundraising emails in its established voice. It supplies its three best-performing appeals as samples. New emails come back closely resembling one of the three, including a specific donor story. What went wrong?","options":[{"text":"The samples were not stripped of specific content and the prompt did not say to match voice while reusing no facts or phrasing.","correct":true},{"text":"Three samples is too few to establish a voice, so the model imitated the nearest one.","correct":false,"tempting":"More samples would dilute the copying but not address the missing instruction; the specific story was reproduced because nothing forbade it."},{"text":"The samples should have been from other charities, to give the model a broader sense of the genre.","correct":false,"tempting":"Would replace their own voice with a generic one, defeating the purpose."},{"text":"Fundraising appeals are too formulaic for voice matching to add anything.","correct":false,"tempting":"Dismisses the technique rather than identifying the missing constraint in this prompt."}]},"prev":{"slug":"showing-what-not-to-do","title":"Negative examples, and when they backfire","url":"https://addaly.com/learn/prompting/showing-what-not-to-do"},"next":{"slug":"examples-from-your-own-archive","title":"Your best examples already exist","url":"https://addaly.com/learn/prompting/examples-from-your-own-archive"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Capturing a voice from three samples', in Getting Real Work out of a Model. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/prompting/capturing-a-voice (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Capturing a voice from three samples","courseTitle":"Getting Real Work out of a Model","module":{"index":3,"slug":"teaching-by-example","title":"Teaching by example"},"answer":null,"keyIdea":"Style transfers through real samples plus named mechanics (sentence length, banned openings, contractions) not through adjectives, and it drifts back to generic over long outputs.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"adjectives-do-not-carry-style","runs":[{"text":"Adjectives do not carry style"}]},{"t":"p","runs":[{"text":"“Professional but friendly.” “Warm, clear, human.” “Confident without being arrogant.” Every organisation has written these words into a style guide, and none of them constrain a sentence. Two writers reading the same three adjectives produce completely different prose."}]},{"t":"p","runs":[{"text":"Style is one of the things prose is worst at describing and examples are best at transmitting. Three real pieces of your own writing will do more than a page of guidelines."}]},{"t":"h2","id":"choosing-the-three","runs":[{"text":"Choosing the three"}]},{"t":"p","runs":[{"text":"Real, not aspirational.","b":true},{"text":" The email you actually sent, the post that actually went out. Not the one you wish you had written, and definitely not something you admire from another company."}]},{"t":"p","runs":[{"text":"Same genre as the target.","b":true},{"text":" Emails teach emails. A blog post will not teach you an email; the shapes are too different and you will get a blog post in an inbox."}]},{"t":"p","runs":[{"text":"Varied within the genre.","b":true},{"text":" One good-news email, one difficult one, one routine one. Three cheerful examples teach cheerfulness, which is not your voice, only your voice on good days."}]},{"t":"h2","id":"strip-what-you-do-not-want-copied","runs":[{"text":"Strip what you do not want copied"}]},{"t":"p","runs":[{"text":"This is the step people skip. Your three samples contain names, dates, projects, and specific facts, and some of those will come through into the new text. Replace with placeholders: "},{"text":"[CLIENT]","code":true},{"text":", "},{"text":"[DATE]","code":true},{"text":", "},{"text":"[PROJECT]","code":true},{"text":". Also cut any signature block, unless you want it reproduced every time."}]},{"t":"p","runs":[{"text":"Then wrap them, so they cannot be mistaken for instructions:"}]},{"t":"code","lang":"","label":"Text","text":"<sample1>...</sample1>\n<sample2>...</sample2>\n<sample3>...</sample3>\n\nWrite a new email in the same voice as the samples above.\nMatch: sentence length, level of formality, how much context is\ngiven before the ask, and whether contractions are used.\nDo not reuse any facts, names or phrasing from the samples.\nSubject: [your actual task]"},{"t":"p","runs":[{"text":"The list of things to match is doing real work. Without it, “same voice” is interpreted broadly, and you get a competent pastiche. With it, the model has four concrete dimensions to attend to."}]},{"t":"h2","id":"naming-the-mechanics-yourself","runs":[{"text":"Naming the mechanics yourself"}]},{"t":"p","runs":[{"text":"You can help enormously by describing the two or three mechanical habits that actually define your voice. Not adjectives: mechanics:"}]},{"t":"ul","items":[[{"text":"Sentences mostly under fifteen words."}],[{"text":"Never opens with “I hope this finds you well”."}],[{"text":"Contractions in email, none in documents."}],[{"text":"One idea per paragraph, no paragraph over four lines."}],[{"text":"British spelling, Oxford comma, no em dashes."}],[{"text":"Says “we” for the company and “I” for a decision."}]]},{"t":"p","runs":[{"text":"Six lines, all checkable. Combined with three samples, this is as close as you will get to reproducible voice."}]},{"t":"h2","id":"the-honest-limits","runs":[{"text":"The honest limits"}],"role":"limit"},{"t":"p","runs":[{"text":"It will drift over a long output.","b":true},{"text":" Voice holds for a paragraph or two and loosens after that. For anything long, generate in sections."}]},{"t":"p","runs":[{"text":"It cannot invent the voice you have not written.","b":true},{"text":" If you have never written a difficult message, three cheerful samples will not tell it how you would."}]},{"t":"p","runs":[{"text":"Generic pull is real.","b":true},{"text":" Model outputs have a centre of gravity (a slightly formal, hedge-heavy, list-loving register) and long generations drift towards it. This is why the banned-words list is worth keeping alongside the samples."}]},{"t":"p","runs":[{"text":"Matching a voice is not matching a judgement.","b":true},{"text":" The samples teach how you say things. They do not teach what you would decide to say, which is why voice-matched drafts often read correctly and land wrongly: right register, wrong emphasis, a concession you would never have made. Read for content first and style second, in that order, every time."}]},{"t":"p","runs":[{"text":"A voice is not authorship.","b":true},{"text":" Producing text in your style does not make it yours in any sense that matters to a reader who thinks you wrote it. Where it matters (a condolence note, a reference, a personal apology) write it yourself. This is not a technical limitation and no prompt fixes it."}]},{"t":"h2","id":"the-free-path","runs":[{"text":"The free path"}],"role":"free"},{"t":"p","runs":[{"text":"None of this needs a paid tool. Keep the three samples and the six mechanical rules in a plain text file. Paste them when you need them, or put them in a project or custom instruction if your tool supports it."}]},{"t":"p","runs":[{"text":"The file is the asset. It took you twenty minutes and it encodes something no template from the internet can, which is how you actually write."}]},{"t":"h2","id":"a-note-for-teams","runs":[{"text":"A note for teams"}]},{"t":"p","runs":[{"text":"The same file works for a team, and it does something a style guide never has: it settles arguments by demonstration. “We write like this” becomes three emails everybody has read rather than an adjective everybody interprets differently."}]},{"t":"p","runs":[{"text":"Where it gets uncomfortable, and where it is worth being honest, is that a shared voice file makes one person’s writing the standard. Choose the samples in the open, with the people whose writing is in them, and revisit the set when the team changes. A voice imposed by a file nobody agreed to is a style guide with extra steps."}]}],"check":{"question":[{"text":"A charity wants fundraising emails in its established voice. It supplies its three best-performing appeals as samples. New emails come back closely resembling one of the three, including a specific donor story. What went wrong?"}],"options":[{"text":[{"text":"The samples were not stripped of specific content and the prompt did not say to match voice while reusing no facts or phrasing."}],"correct":true,"tempting":null},{"text":[{"text":"Three samples is too few to establish a voice, so the model imitated the nearest one."}],"correct":false,"tempting":[{"text":"More samples would dilute the copying but not address the missing instruction; the specific story was reproduced because nothing forbade it."}]},{"text":[{"text":"The samples should have been from other charities, to give the model a broader sense of the genre."}],"correct":false,"tempting":[{"text":"Would replace their own voice with a generic one, defeating the purpose."}]},{"text":[{"text":"Fundraising appeals are too formulaic for voice matching to add anything."}],"correct":false,"tempting":[{"text":"Dismisses the technique rather than identifying the missing constraint in this prompt."}]}],"order":[1,0,3,2],"why":null},"prev":{"slug":"showing-what-not-to-do","title":"Negative examples, and when they backfire","minutes":8},"next":{"slug":"examples-from-your-own-archive","title":"Your best examples already exist","minutes":8}}}