How do you sound better over text?

The reply box has been open for ten minutes and holds one deleted line. It could be an apology, a first message to someone you met on Saturday, a reply to a vague “hey”, or the message you have now left for three weeks. I co-founded Subtext, and the drafts I see stall most often are not hard because of the words. They are hard because the person writing has not decided what the message is for. If someone has had bad news, what to say when you don’t know what to say is the piece you want. This one covers everything else. Further down, a real run of Subtext reads a first message to someone met on Saturday that hedges its invitation twice, with the one tag and the three versions it came back with.

Two things the research supports change where to put the effort. You tend to overestimate how clearly your tone comes across in writing, at least with sarcasm, humour and emotion, because you hear the tone you meant and the reader does not¹. And polish buys less than it seems to. Long words used where short ones would do made writers seem less intelligent, not more². In two experiments on comforting someone after a parent’s death, a moderately supportive message was rated about as effective as a highly elaborated one³. That is grief, not everyday texting, but it points the same way, that past a decent floor, more craft buys little. So the work is to get the purpose right, clear the floor, and then stop.

You are stuck on a decision, not a sentence

Ask someone what they are trying to say and they usually tell you what they are trying not to say. Not needy. Not cold. Not like they are making excuses. That is a list of things to avoid, and you cannot write a message out of avoidances.

The freezing itself is ordinary. A meta-analysis of 50 experiments found that, on average, more options did not overwhelm people⁴. A later one found overload more likely when the decision is hard, the options are complex, you are unsure what you want, or you just want it over with⁵. A blank message plausibly fits the third. You are not paralysed by having too many words available. You are paralysed because you have not decided what the message has to do, so every draft is judged against a target that does not exist yet. I took the deciding part apart in why replying to a simple message feels like so much work, and the choice research in how to stop overthinking a text.

One more thing, because the word anxiety gets attached to all of this. Hesitating over a message, rewriting it five times and worrying how it will land are normal, and none of them is a diagnosis. Social anxiety disorder is a clinical condition with real impairment that lasts, and the US National Institute of Mental Health is explicit that occasional stress and anxiety are normal⁶. If the fear is persistent and getting in the way of your life, that is a conversation for a professional, not an article. For everyone else, the blank screen is a decision problem, and decision problems have a procedure.

So decide first. Almost every text you will struggle with is doing one of nine jobs.

• Comfort. Something hard happened to them.
• Repair. You caused hurt or dropped a ball.
• Calm things down. There is tension and you are still upset.
• Open. You are starting something with someone new.
• Keep going. They sent something vague and you want the conversation to continue.
• Come back. Your reply is overdue.
• Ask. You need something, at work or with plans.
• Decline. You are saying no, or you need to stop for now.
• Acknowledge. You cannot decide yet, and silence is becoming the message.

Pick one. If the message is doing two, pick the one that matters more and give the other a single sentence. Once you know the job, the wording problem gets small.

Three moves, in order

Name the thing. Say one true sentence. Then ask one easy question, or close. That is the whole method.

Name the thing. Say what you are responding to: the interview, last night, the message you left too long. In five studies, Packard and Berger found customers were more satisfied when the employee replying used concrete words, which they read as a sign of listening⁷. “Good luck with the driving test on Friday” beats “good luck with everything”. It is the cheapest move in the list and the one people skip when they are nervous, because a specific noun feels like committing. What makes a reply land well has the numbers.

One true sentence. Your actual position or feeling, in plain words, without a cushion of qualifiers. Plain is usually enough, which is what the grief study above suggests³. In three of Oppenheimer’s five experiments, making the wording more complex lowered how intelligent readers judged the writer, and two more showed the effect runs through how hard the text is to read². Those were essays, not texts, and the outcome was intelligence, not warmth, so I would take only this much from it, that harder wording made readers think less of the writer. “True” is the important word. A tool can help you find the sentence. It should not invent a feeling, a memory or a promise you do not have.

One question, or a close. Which one depends on the job. People who ask questions about what the other person just said tend to be better liked in live conversation, though the paper behind that is disputed in parts, which I cover in the reply article⁸. When someone is overwhelmed, a question is work. “No need to reply” can be the kinder close, and there is no rule that every text needs a question on the end.

And if you cannot do any of that yet, acknowledge. “Got this, I’ll reply properly tonight” turns silence into a signal, on one condition. No study has tested holding replies directly, but the nearest one sets it. Lew and colleagues showed 131 US adults versions of a customer service chat. Generic replies did clearly worse than specific ones, and fast generic replies had the lowest scores of all⁹. A holding reply that names nothing looks like that worst combination. “I saw the plan for Mum’s birthday and I’ll answer about the restaurant tomorrow” names something, and it takes the same few seconds to send.

Polish comes after the three moves exist, not before. Matching their length, cutting the inflated words and checking the full stop all belong to a second pass on a message that already does its job.

What it looks like, situation by situation

None of these is the correct text. Each is one way through, with the job named and the move that changed it. Acknowledging is covered above, comfort has its own article, and the last example, checking what they meant, is a move that fits several jobs.

Repair. You snapped at someone last night.

Before “Sorry if you were upset by what I said, I was just stressed. Can you forgive me?”

After “I was sharp with you last night and that wasn’t fair. I’m sorry. I should have said I needed a break. Can I call you later?”

“Sorry if” moves the fault onto their reaction, and “forgive me?” asks them to comfort you. The after names what you did, owns it, says what you would do differently, and offers a call. The research on written apologies is in how to apologise over text.

Calm things down. The fight is over and you are still hurt.

Before “Fine. Whatever you think is best.”

After “I’m still a bit hurt, but I don’t want this to sit between us. Can we talk tomorrow when we’re both calmer?”

“Fine.” on its own reads as a door closing, whatever the full stop adds. One true feeling and a time to talk keeps it true without reopening the argument, which is the mistake what to text after an argument is about.

Open. Someone you met at a workshop on Saturday.

Before “Hey! It was really nice to meet you!!! We should totally hang out sometime!!”

After “Hi, it’s Jules from Saturday’s pottery workshop. Your glaze tip saved my bowl. Are you going to the next session?”

Who you are, one specific shared detail, one question that is easy to answer. The before has none of the three, and six exclamation marks doing their work instead.

Keep going. They sent “hey”.

Before “Hey.”

After “Hey! Good to hear from you. How did the half marathon go?”

Keep it short, add one degree of warmth, give them something to answer. Mirroring a flat reply with your own flat reply is how conversations end without anyone deciding to end them.

Come back. You left their message for three weeks.

Before “Hiii so sorry, life has been insane lately lol”

After “Sorry I went quiet for three weeks. I did read your message about the move, and I want to hear how it went. Are you free for a call this week?”

One apology, then proof you read theirs, then a question that turns the attention back to them. Why this message gets harder to send the longer you leave it is in why you can’t reply to messages you actually want to answer.

Ask. An invoice nobody has answered.

Before “Just circling back per my last email as I haven’t heard anything.”

After “Hi Alex, following up on the invoice from 3 October. Could you let me know when it’s likely to be paid? Thanks.”

A specific reference, one clear ask, plain words. “Circling back” and “as I haven’t heard anything” are stock phrases that get read as a dig whether or not you meant one, which is the subject of the passive-aggressive email.

Decline. An invitation you are not going to.

Before “I can’t come because I’ve had a really long week and big groups are hard and I have to get up early and I’m sorry and I hope you’re not mad…”

After “Thanks so much for inviting me. I’m going to sit this one out, but I’d love to get coffee next week. Does Tuesday work?”

A reason is optional. A clear no plus an alternative avoids both the long explanation and the “maybe next time” that keeps the request alive. The work version is in how to say no to extra work.

Check, don’t guess. They replied “fine” and you cannot tell what it means.

Before “Fine?”

After “I’m not sure how to read ‘fine’ here. Do you mean the plan works for you, or would you rather change it?”

Saying what you cannot tell and asking what they meant replaces an hour of rereading. How much a short message actually carries is in what does this text mean?.

The comfort job is missing from this list on purpose. When the news is bad, the floor matters more than anywhere else, and the evidence is different enough that it has its own article.

What “sounding better” actually means

People who ask me how to sound better usually mean one of two things, and they pull in opposite directions. Some want to sound warmer and less abrupt. Some want to sound less needy and more composed. Both groups reach for the same fix, which is more words, and more words is the one change the research does not support.

Longer is not warmer. Kemp, Kovacic and Beyersmann had 200 Australian undergraduates rate exchanges between friends of one to eight words¹⁰. Longer messages drew slightly less negative ratings, and at six to eight words the full stop no longer made a significant difference. Nothing longer was tested, so it says nothing in favour of a six line reply to “Thursday works?”. That reply is not warmer than “Yes, see you then”. It makes the reader search for your yes.

More formal is not better either. The long words finding points the same way, and it matters for second language writers, who were often taught formal English as the safe choice. “I would like to kindly inform you that the music is being very loud” is correct and reads as cold. “Hi, it’s Ana from flat 4. The music late at night is keeping me awake. Could you turn it down after 11?” is shorter and lands. I’d flag that Kemp’s study included only people whose first language was English, so its conventions have not been shown to hold for second language writers. Do my English texts sound polite? covers what a non-native writer can and cannot check.

The full stop is real, small and specific. Gunraj and colleagues had undergraduates rate one word replies such as “Okay” and “Sure”, and the version with a full stop read as less sincere in texts, though not in handwritten notes¹¹. It is a finding about short informal replies, not a rule that full stops are rude, and the claim you will see beside it, that an exclamation mark makes a reply more sincere, came from later, unpublished work. The whole story is in does a full stop make a text sound angry?. On a short reply where you mean warmth, add a word before you worry about the dot; “Sure, sounds good” does more than dropping the full stop. Several exclamation marks read as nerves, and none belong in a message about grief or a fight.

Emoji work when they match. In Huh’s experiment with 260 US adults, an emoji that fitted the text raised how responsive the sender seemed, in exchanges with a close friend¹². One study, hypothetical exchanges, emoji chosen to match. It does not show that an emoji makes a serious message warmer. Follow the other person’s usage and you will not be far wrong.

Match their style, up to a point. Ireland and colleagues found that similarity in the small, unconscious parts of language, the pronouns and prepositions, predicted whether relationships started and lasted¹³. That is correlational. I found no study testing whether deliberately copying someone’s slang or emoji in texts makes them like you more, and I think it reads as mimicry. Match their length and their formality. Leave their vocabulary alone.

Plain words. This is the one that generalises. Cut the synonym you reached for because it sounded more serious. Cut the qualifier in front of the feeling. Cut “just”. If the message still says what you mean, it was padding.

None of these cues matters as much as the move it decorates. A well punctuated message that skips the specific noun and dodges the true sentence is still a message that does not do its job.

Subtext reads a draft for some of this list. It marks words that could read as harsh, cold, passive aggressive, vague or over-apologetic and says why, and what you change stays your decision.

When a tool helps

AI help makes drafting faster and, with standard suggestions, the words more positive, and it costs you something when the other person suspects it. In Hohenstein and colleagues’ first experiment, people who believed their partner was using smart replies rated that partner as less cooperative, while partners who actually used them were rated more cooperative¹⁴. In a separate study of 34 student pairs, smart reply users were not rated less warm or competent¹⁵. Help is not obviously punished; a reply that reads as outsourced is. I go through those studies in what makes a reply land well, and what a full AI rewrite does to your voice in does AI make your texts sound like a robot?.

As the co-founder of a tool in this category, I take that evidence as a line between two jobs. Helping you get unstuck and giving you a second read on tone is the useful one. Replacing your message with the tool’s is the damaging one. Subtext is built around the first. It starts from whatever you have, a half finished draft, a few words or a voice note, and when the job is not clear it asks what you want to get across before it writes anything. Then it writes up to three versions close to your words, because the choice between warmer and more direct is the decision this article is about and it is yours to make. Each version is marked safe to send or heads-up, and you can ask for any of them shorter, warmer, more casual or more confident. It shows how much it changed from what you wrote, and it never sends anything. It cannot tell you what the other person meant, and the site’s FAQ says so. Nor can it make anyone reply.

Here is that second job on a first message, made up for the run, to someone met on Saturday, ending “we should do it again sometime maybe? Only if you want to obviously, no pressure at all haha”. Subtext named the feeling as Hopeful and tagged one thing, Mixed messages, with the note that the maybe and the “no pressure” together sound unsure and make the invite easy to dodge. It wrote three versions, and all three were marked safe to send. The first, Close to yours, turned the maybe into a question, “Would you like to do it again sometime?”, and kept the easy tone. The third, Confident rewrite, went furthest and proposed “Want to grab a coffee or a drink this week?”, which the draft never said, and Subtext said so under the versions so you can change it. That is the app’s real output, unedited. It did the middle move for the sender, one clear question in place of two hedges, and left the choice of how much to propose with them.

Try Subtext in your browserDownload on the App StoreGet it on Google PlayGet the app

Two things I would do with any tool’s output, mine included. Put one detail back in that the tool could not have known, because that detail is both what makes the message yours and what makes it read as yours. And if the message is an apology, a condolence or a love note, the sentence in the middle should be one you wrote. The wider comparison of what rewriters do to your words is in do AI message rewriters actually work?.

What to paste, and when to say you had help

Pasting a conversation into any AI tool shares another person’s words with a service they never agreed to. I think about this a lot, partly because people paste screenshots into Subtext every day.

The principle I’d use is data minimisation, which the GDPR states as using only the information a task needs¹⁶. Whether the regulation applies to your personal use of a tool is a separate legal question, and this is a design rule, not legal advice. So paste the lines you need and not the thread, crop names and numbers out of screenshots, and keep anything about the other person’s health, money or private life out of a tool whose only job is to help you phrase a reply. Which tools keep your text and which train on it changes by the month, and I read the policies in which AI writing assistants train on your messages.

On disclosure, there is no rule and I would not invent one. Nobody expects a label on a text you ran through a spell check. In thirteen experiments by Schilke and Reimann, people who disclosed using AI were trusted less, and in the experiments that tested it, being exposed by someone else cost more trust than saying so yourself¹⁷. In a 2026 study of messages between partners, admitting AI help on a lightly edited apology made it land better, and admitting it on a fully AI drafted one made it land worse¹⁸. So the more a message claims to carry your own feeling, responsibility or memory, and the more of it the tool wrote, the stronger the case for rewriting it in your own words or saying you had help. “I used a writing tool to get this in order, but this is what I mean” keeps the ownership where it belongs. The studies did not test denial, but being exposed was the worst outcome they measured, so if someone asks you directly, do not deny it.

Advice with nothing behind it

“Never start with hey.” I found no study that compares openers. “Hey” is not bad, it is low on information, and it is fine between people who always start that way.

“Always end with a question.” The question asking research is about live conversation between strangers getting to know each other⁸. In grief or exhaustion, a question is a task.

“A longer message shows you care.” The closest findings do not support it. A highly elaborated supportive message was rated about as effective as a moderate one³, and the one study that varied length stopped at eight words¹⁰.

The timing rules (“wait three hours”) and the texting anxiety statistics that circulate without a study behind them are taken apart in what makes a reply land well and why texting gives you anxiety.

The order to do it in

Decide what the message is for before you type. If you cannot, send the acknowledgement, and make it name something.

Then three moves. Name the thing. Say one true sentence, in the plainest words that still mean it. Ask one easy question if you want the conversation to go on, or close if the other person needs the pressure off.

Only then look at the surface. Match their length, not their vocabulary. On a short reply you mean warmly, add a word before you worry about the full stop. Cut the synonym and the qualifier. If a tool helped, put one detail back that it could not have known, and if the message is supposed to carry your own feeling, make sure the middle sentence is yours.

The thing you were agonising over was almost never the words. It was the decision. Make it, write the smallest true message that does the job, and send it before you have time to make it longer.

Sources

Numbered in the order they appear above. Where a figure could not be verified against the primary source, the text says so. Checked 24 October 2026.

1. Kruger, J., Epley, N., Parker, J., and Ng, Z.-W. (2005). Egocentrism Over E-Mail: Can We Communicate as Well as We Think? Journal of Personality and Social Psychology, 89(6), 925 to 936. Five experiments, US undergraduates, email.
2. Oppenheimer, D. M. (2006). Consequences of Erudite Vernacular Utilized Irrespective of Necessity: Problems with Using Long Words Needlessly. Applied Cognitive Psychology, 20(2), 139 to 156. Five experiments, three of them manipulating how complex a text’s wording was; the outcome was perceived intelligence, not warmth.
3. Tian, X., and Solomon, D. H. (2025). The Role of Person-Centered Messages, Parallel Disclosures, and Reactance When Communicating Support for Parental Death. Communication Research, 52(8), 1116 to 1144. Two experiments comparing low, moderate and high person-centred messages.
4. Scheibehenne, B., Greifeneder, R., and Todd, P. M. (2010). Can There Ever Be Too Many Options? A Meta-Analytic Review of Choice Overload. Journal of Consumer Research, 37(3), 409 to 425. 63 conditions from 50 experiments, N = 5,036; mean effect virtually zero, with wide variation.
5. Chernev, A., Böckenholt, U., and Goodman, J. (2015). Choice overload: A conceptual review and meta-analysis. Journal of Consumer Psychology, 25(2), 333 to 358. 99 observations, N = 7,202; four moderators. Follows Scheibehenne, Greifeneder and Todd (2010), listed separately above. A corrigendum correcting statistical reporting appeared in 26(2), 312.
6. US National Institute of Mental Health. Social Anxiety Disorder: More Than Just Shyness. Public information page.
7. Packard, G., and Berger, J. (2021). How Concrete Language Shapes Customer Satisfaction. Journal of Consumer Research, 47(5), 787 to 806. Five studies, including text analysis of over 1,000 real customer calls and emails.
8. Huang, K., Yeomans, M., Brooks, A. W., Minson, J., and Gino, F. (2017). It Doesn’t Hurt to Ask: Question-Asking Increases Liking. Journal of Personality and Social Psychology, 113(3), 430 to 452. Disputed in part by Kluger and Malloy (2019), with the authors’ reply; carries a published correction from March 2025. Observers reading transcripts did not share the preference.
9. Lew, Z., Walther, J. B., Pang, A., and Shin, W. (2018). Interactivity in Online Chat: Conversational Contingency and Response Latency in Computer-Mediated Communication. Journal of Computer-Mediated Communication, 23(4), 201 to 221. 131 US adults watched one of four simulated chats; fast generic replies scored lowest, though not significantly below slow generic ones.
10. Kemp, N., Kovacic, R., and Beyersmann, E. (2025). Is the period really “pissed”? The effect of punctuation and message length on perceptions in digital communication. Telematics and Informatics, 97, 102241. 200 Australian undergraduates, first language English only.
11. Gunraj, D. N., Drumm-Hewitt, A. M., Dashow, E. M., Upadhyay, S. S. N., and Klin, C. M. (2016). Texting insincerely: The role of the period in text messaging. Computers in Human Behavior, 55, 1067 to 1075. The exclamation mark result was described by the university’s release as unpublished follow-up work.
12. Huh, E. (2025). The impact of emojis on perceived responsiveness and relationship satisfaction in text messaging. PLOS ONE, 20(7), e0326189. 260 US adults aged 23 to 67, hypothetical exchanges with a close friend.
13. Ireland, M. E., Slatcher, R. B., Eastwick, P. W., Scissors, L. E., Finkel, E. J., and Pennebaker, J. W. (2011). Language Style Matching Predicts Relationship Initiation and Stability. Psychological Science, 22(1), 39 to 44. Correlational.
14. Hohenstein, J., Kizilcec, R. F., DiFranzo, D., Aghajari, Z., Mieczkowski, H., Levy, K., Naaman, M., Hancock, J. T., and Jung, M. F. (2023). Artificial intelligence in communication impacts language and social relationships. Scientific Reports, 13, 5487. Two randomised experiments, 1,020 participants recruited; pairs chatted live.
15. Mieczkowski, H., Hancock, J. T., Naaman, M., Jung, M. F., and Hohenstein, J. (2021). AI-Mediated Communication: Language Use and Interpersonal Effects in a Referential Communication Task. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1). 34 student pairs, mean age 20. Receivers reported somewhat lower social attraction, not significant after correction.
16. European Commission. Principles of the GDPR.
17. Schilke, O., and Reimann, M. (2025). The Transparency Dilemma: How AI Disclosure Erodes Trust. Organizational Behavior and Human Decision Processes, 188, 104405. Thirteen experiments; exposure by a third party was tested in some of them; denial was not tested.
18. Fan, G., Liu, D., and Pan, L. (2026). Is It Still You? Attributing Authorship and Authenticity in AI-Assisted Romantic Communication. CHI 2026. Randomised experiment, 704 people.