Write better

~ min read

30-second summary

“Writing better” doesn’t mean “letting the AI write for you”: it means using it on your draft, not as a replacement.
Three patterns that keep your voice: your draft, AI’s critique (diagnosis, not rewrite), more versions (you take pieces, not a whole one), emotional temperature (check what the tone projects before sending).
The AI has a default register that’s polite but anonymous. If you copy and paste for weeks, that default becomes your standard voice, and your readers notice.
Not everything fits: coordination emails yes, messages to close people no. Where sincerity is the content, don’t delegate.

You’ve got an email to send. The deadline is in ten minutes, the right tone isn’t coming, and you’re staring at a blank page.

You can ask the AI to write it for you and paste what it gives you. Or you can ask it to help you write it better yourself. They look like the same thing, but they aren’t. The first option returns a text that would work for anyone with that generic problem: polite and correct, but without a signature. The second starts from your own attempt, even a clumsy one, and works on top of it.

This lesson is about the second option. It works better than the first on two counts: the texts that come out feel like yours because they are, and over time you become a sharper reader of your own writing.

Three patterns that keep your voice

1. Your draft, AI’s critique

The most counterintuitive, and the most useful. You write the first version yourself, even badly, fast, without rereading it. The imperfect draft serves exactly that purpose: it forces the AI to react to your ideas instead of generating its own. Then you hand it to the AI to tell you what isn’t working, not to rewrite it.

“Here’s my email. Does it feel too long? Where am I ambiguous? Is there a passage where I slip from polite into passive-aggressive without noticing?”

What you get isn’t a new text, it’s a diagnosis: three or four observations pinned to specific sentences. From there you rewrite yourself, choosing what to keep, what to cut, what to move. The voice stays yours, because the scaffolding stays yours.

A handy variant: ask for critique in separate layers. “First tell me what isn’t working on the level of clarity. Then on tone. Then on length.” Three different diagnoses, and by separating them you see more clearly what to change first. If length isn’t a problem, you don’t waste time on it.

An example. I need to tell a client about a delay on a project, I’ve jotted down a draft that feels bureaucratic to me but I can’t put my finger on the problem. I hand the text to the AI asking only for a diagnosis.

You

I've jotted down this email for a client I need to tell about a delay. Don't rewrite it: tell me in three points what isn't working, then I'll rewrite it myself. "Good morning, I'm writing to inform you that unfortunately we will not be able to deliver the project within the originally planned timeframes due to some operational difficulties we've encountered over the past few weeks. The new delivery date will be May 15. We apologize for the inconvenience and remain available for any clarifications." I've jotted down this email for a client I need to tell about a delay. Don't rewrite it: tell me in three points what isn't working, then I'll rewrite it myself. "Good morning, I'm writing to inform you that unfortunately we will not be able to deliver the project within the originally planned timeframes due to some operational difficulties we've encountered over the past few weeks. The new delivery date will be May 15. We apologize for the inconvenience and remain available for any clarifications."

Three observations. 1. The reason for the delay is vague. "Some operational difficulties" tells the client nothing and lets them imagine something worse than the reality. If the reason is simple (a supplier running late, someone out sick, an external dependency), stating it bluntly builds trust. 2. The new date arrives at the end as an anonymous data point. Put it up front, right after the news: that way the client knows what to expect without having to hunt for it. 3. "We apologize for the inconvenience" is a voicemail corporate format. Replace it with an apology that acknowledges what the delay costs them (a rescheduled launch, a replanning job). Three observations. 1. The reason for the delay is vague. "Some operational difficulties" tells the client nothing and lets them imagine something worse than the reality. If the reason is simple (a supplier running late, someone out sick, an external dependency), stating it bluntly builds trust. 2. The new date arrives at the end as an anonymous data point. Put it up front, right after the news: that way the client knows what to expect without having to hunt for it. 3. "We apologize for the inconvenience" is a voicemail corporate format. Replace it with an apology that acknowledges what the delay costs them (a rescheduled launch, a replanning job).

It doesn’t hand you a finished text, it hands you the scalpel. The three pinned observations are the map of what you’ll go change in rewriting.

2. More versions, none entire

The second pattern flips the typical way of using the AI. Instead of asking for the right version, ask for three or four calibrated differently, and don’t take any of them in full.

“Give me three versions of this email to ask for a refund. One direct and dry, one diplomatic, one formal. Three or four sentences each, no more.”

The specific labels matter: asking for “three versions” without more gives you three often-similar variants. Different names (“direct”, “diplomatic”, “formal”) force the AI to move the register in recognizable directions.

The value isn’t picking the best one. It’s seeing how the same thing changes shape, and figuring out which register sounds like yours. Usually it’s none of the three: it’s a hybrid, the opening from the first version, the middle passage from the second, a closing line from the third.

Recognizing your own voice isn’t automatic, especially if you write a lot of work email and may have picked up some AI tics already. Two practical clues. Compare with an old text of yours you liked (an email to a friend, a diary note, anything): there your voice was more naked. Then read the three AI versions out loud and flag the spot where the sentence slips by without catching: that’s the passage that comes closest to how you talk.

Instead of copying, compose. Take the opening from one version, an argument from another, a close from a third. Then reread and smooth the transitions. What comes out is a text that wasn’t in any of the three, but that the AI helped you assemble. The mosaic is worth it when none of the three versions convinces you in full. If one is already close to good, retouching one sentence is faster and leaves fewer seams.

More ambitious variant: paste two or three past texts of yours you’re happy with and ask “pick up the cadence of these texts, not their content”. It isn’t magic, but the AI grabs some of your rhythm tics (where you break sentences, how long you extend subordinates) and carries them into the new versions.

3. Emotional temperature

The third pattern is where the AI is unexpectedly good: reading the emotional tone of a text. The explanation isn’t magic: it has read millions of texts where people explicitly commented on the tone they received (“sounds cold to me”, “it’s passive-aggressive”), and has learned which surface signals trigger those reactions. It can tell when a “certainly” sounds ironic, when a formal close lands distant, when a slightly off-key adjective shifts the whole message.

“Read this email and tell me what tone it’s putting across. Does it sound cold? Passive-aggressive? I want to say no to a coworker without making them feel brushed off.”

Particularly useful in high-stakes cases: saying no to someone who asked you for something, negotiating a raise, responding to feedback that stung, delivering bad news. There the tone is the message, more than the words.

The pattern holds in other languages too: the AI has read enough text in English, Spanish, French to read their register. One caveat: it sometimes projects the register norms of another culture onto yours (a “cold” in American English isn’t a cold in British English or Italian). Read the diagnosis with the context shift in mind.

You

I want to say no to Marco, a coworker, without making him feel brushed off. Read this email and tell me what tone it puts across: "Hi Marco, thanks for thinking of me, but at the moment I can't take on this project. I hope you understand. Let me know if you need anything else. Luca"

It sounds polite, but a bit distant. Two passages work against your intention. "I hope you understand" right after a refusal can read as an early shutdown, as if you didn't want to discuss it. "Let me know if you need anything else" at the end of a no is boilerplate: what else would Marco need? You've just told him you're not taking the project. If Marco is a coworker you work with often, give the real reason for the no and drop the generic close. For example: "Thanks for thinking of me. I'm buried in deadlines right now and wouldn't be able to give the project the attention it deserves. If you need a hand in June, give me a shout." It sounds polite, but a bit distant. Two passages work against your intention. "I hope you understand" right after a refusal can read as an early shutdown, as if you didn't want to discuss it. "Let me know if you need anything else" at the end of a no is boilerplate: what else would Marco need? You've just told him you're not taking the project. If Marco is a coworker you work with often, give the real reason for the no and drop the generic close. For example: "Thanks for thinking of me. I'm buried in deadlines right now and wouldn't be able to give the project the attention it deserves. If you need a hand in June, give me a shout."

The diagnosis is useful even when you don’t apply the rewrite it proposes: it flags things you’d miss rereading on your own.

The diagnosis can also be wrong: the AI sometimes projects exaggerated readings onto neutral texts. If a reading surprises you drastically, ask “what in the text are you seeing this in?” and read the passages it cites. If those passages don’t hold up, throw out the diagnosis; if they do, it has a point even if your gut said otherwise.

Difference from pattern 1: there you ask for a technical critique (clarity, structure, length); here you ask for a reading of the emotional register. Same muscles, different focus. In delicate cases you can combine them: first the technical diagnosis, then the tone one, separately, so you see whether an email is unclear and cold together or just one of the two.

The trap of the average voice

The AI has read millions of work texts, and has learned how they sound on average. Its default is a polite but anonymous register: correct, smooth, interchangeable. That’s AI-polish, the impersonal glaze of generated text, readable but unsigned.

The problem starts when you copy its versions without working them. For a week it doesn’t show; after a month your emails all start with “Thank you for…” and close with “Please let me know if…”, and look like they came from the same mold. Your readers can’t name it, but they can feel it.

A concrete example. Your quick draft, before running it through the AI:

“Hi Marco, the offer’s interesting but the timing doesn’t work. I’ve got too much on my plate until mid-June. If you’re up for it, let’s pick this up in July?”

The version coming out of the AI when you ask it to “improve” the text without specific instructions:

“Hi Marco, thank you for presenting me with this interesting opportunity. Unfortunately, at the moment my schedule is particularly full until mid-June and I’m unable to dedicate to this project the attention it would deserve. I remain available to pick up the conversation from July onward, if that works for you.”

The second is longer, more formal, and says exactly what the first says. The first is more yours: clear, calibrated to your relationship with Marco, no padding. If you make this swap ten times a day for a month, the register of the second becomes your standard.

Practical counter-move: read the final version out loud before sending it. If it sounds like a call-center reply, cut. A second test: try to say in two words what you want to get across. If the summary is short and clear but your written version is three times as long, there’s padding. The most common padding: long connectors (“unfortunately at the moment”), redundant politeness formulas, filler adverbs (“particularly”, “effectively”, “certainly”).

When to use it, when not to

Three buckets, not symmetric. Most emails live in the middle, and calibration is case by case. The map below is for tuning yourself, not boxing yourself in.

Use the AI freely for short coordination emails (confirmations, follow-ups, updates), internal documents where clarity matters and voice doesn’t (memos, minutes, operating instructions), and technical texts where the content does all the work. Here the AI’s average voice is adequate, and the time saved is real.

Use it carefully, rereading twice and putting your hand back in, for texts where you state a position of yours (reviews, open letters, posts that carry your name), delicate emails (refusals, negotiations, constructive critiques), and communications to people who know how you usually write. The caution here isn’t on using the AI itself, it’s on the padding: an AI-polish email arriving from someone who usually writes dry sounds like a different person, and the recipient registers it without being able to say so.

Don’t use it to write from scratch: messages to people close to you (friends, family, partner), intimate texts (notes, cards, private letters), communications where sincerity is the content (apologies, condolences, personal declarations). The issue isn’t ethical in the abstract: it’s that the human on the other side often notices, even without being able to name it. The AI can help you reread, smooth a harsh passage, check whether the tone is what you meant (that’s pattern 3, applied in small). Not speak on your behalf.

What comes next

So far you’ve worked with text that you write or paste. Modern AI goes further: it reads photos, screenshots, PDFs, tables. The next lesson, Photos, images, and files, enters that territory.

Check what you understood

About Privacy Open manual, written by a real person, in collaboration with AI.