How do I check a message before sending it?

Write the draft but don't send it. Paste it into an AI with a prompt that asks how it comes across, which lines are risky, and for a calmer rewrite in two tones (polite and firm) then pick the version that keeps your point without the harsh edge.

Can AI rewrite my angry email?

Yes, and it's useful for exactly this, but treat it as a second reader rather than an oracle. Models are measurably weaker on sarcasm, particularly on the borderline cases humans themselves disagree about, and they tend to adopt the sender's perspective, which is the same bias you're trying to correct for. They also carry documented biases around dialects and language varieties, so a non-native English writer may be told a perfectly normal sentence sounds wrong. Remove names and identifying details first, and treat the output as a draft to review, not a final word on anything with legal weight.

What should I remove before pasting a work message?

Real names, company and client names, invoice numbers, addresses, and account details. The tone fix doesn't need them — describe the situation generically (for example, "a client refuses to pay an invoice"), and paste only the message itself.

Should I ask more than one AI to check the same draft?

It helps, and you don't have to do it by hand. Secret Chat's AI Council sends the draft to several models from different companies at once, shows each read as that model finishes, and then has a referee mark which model flagged each problem, which disagreed, and which never raised it, so a line three models independently call threatening is settled, and a line they split on is the one you decide yourself. A two-model Duet is free within the daily quota; credits buy the ChatGPT/Claude/Gemini/Grok panels and the larger Trio and Quartet. Two caveats: agreement is evidence, not proof, models share training data and can share the same tone and dialect biases, and a council sends your draft to every model on the panel plus the referee, so strip names and identifying details before you paste it.

Why use Secret Chat instead of a normal AI chat?

Because checking a tense draft in a normal chat leaves your most emotional drafts in a saved history. Secret Chat is private by default — local storage, training opted out, identity anonymized, automatic deletion requests, and a PDF privacy report — so the draft doesn't outlive the moment. The chosen model still processes your prompt to answer it.

How to Check a Risky Work Message Before You Send It

Published June 17, 2026 · Updated August 19, 2026

A great deal of damage at work is done not by big decisions but by one message sent thirty seconds too early: the angry reply to a client, the sharp line in a payment dispute, the defensive "that's not my fault" to a colleague. You meant to stand your ground; it came out sounding rude, threatening, or weak. And once it's sent, you can't un-send it.

The fix is a small habit: check the message before it reaches another human. This guide shows how to read a draft for risk, rewrite it so it keeps your point without the line you would regret, and do all of that privately, so your half-written emotional drafts don't pile up in someone's chat history. Near the end we look at Secret Chat, a private AI app built around one simple move: drop your draft here first.

Why "Send Later" Beats "Send Now"

When you're annoyed, your draft is written for you — to vent, to be right, to feel better for a moment. The person receiving it reads something else entirely: tone, intent, and threat level. The gap between those two is where working relationships get damaged.

There is good evidence that this gap is real and that we cannot see it ourselves. In a set of five experiments published in the Journal of Personality and Social Psychology, Kruger, Epley, Parker and Ng had people send statements in a deliberate tone (sarcastic, serious, angry or sad), and predict how well the recipient would read it. Senders were around 90% confident their sarcasm would land. Recipients who heard the statements identified the tone about three-quarters of the time; recipients who read them by email performed at a rate indistinguishable from chance. The authors put it down to egocentrism: you hear your own intonation as you type, and can't detach from it to read the words as they will actually arrive.

A short pause closes that gap. You don't need an hour or a mediator; you need a second pair of eyes that reads the message the way the recipient will, points out the risky parts, and offers a calmer version you can actually send. That's a task an AI is well suited tt, as long as you do it privately.

The Workflow: Check, Understand, Rewrite

The habit is short enough to run on any message that makes your jaw tighten:

Don't send the emotional draft immediately. Write it, then stop.
Paste it into a private message-checking tool.
Get a read on the risk — what sounds rude, threatening, weak, or unclear.
Understand why it lands badly, not just that it does, so you learn the pattern.
Generate a safer rewrite that keeps your actual point.
Pick the variant that fits: polite, firm, shorter, or client-friendly.

If you want a prompt you can reuse every time, paste this above your draft before you send anything:

Before I send this message, check it for me. Tell me how it might come across to the person receiving it — does it sound rude, angry, threatening, weak, passive-aggressive, or unclear? Point out the riskiest lines. Then rewrite it to be calm, clear, and professional while keeping my main point. Give me two versions: one polite and one firm. Here's my draft:

That single prompt turns a vent into a decision: you see the risk, you see two ways to handle it, and you choose — instead of letting the first angry draft choose for you.

Keep the Details Out of It

A draft-checking tool only needs the wording, not the dossier. Before you paste, strip out anything that doesn't need to be there:

Remove real names — yours, the client's, colleagues'.
Replace company and client names with placeholders like "the client" or "the vendor."
Drop invoice numbers, addresses, and account details.
Paste the message itself, not the whole email thread or contract.

For example, instead of:

"John Smith from ACME Ltd refuses to pay invoice #123…"

Use:

"A client refuses to pay an invoice after delivery. Help me write a firm but professional reply."

You get most of the benefit with far less exposure, since the tone fix rarely depends on the names, so leave them out. Two honest caveats. Stripping the surrounding thread can cost the model context it needed, so a rewrite of an anonymised fragment is sometimes blander than one that saw the whole exchange. And a genuinely distinctive dispute can stay recognisable even with every name removed.

Examples: Risky Drafts and Safer Rewrites

Three situations most of us recognize.

Example 1: Angry Client Email

Original: "This is ridiculous. You keep changing requirements and wasting my time."

Safer rewrite: "I want to clarify the scope before continuing, because the recent changes are affecting the timeline and cost. Could we agree on the final requirements before the next step?"

Example 2: Payment Dispute

Original: "You still haven't paid me. This is not acceptable."

Safer rewrite: "I wanted to follow up on the unpaid invoice. Could you confirm when payment will be processed? If there is an issue, I'm happy to clarify it."

Example 3: Internal Work Conflict

Original: "Nobody told me about this, so it's not my fault."

Safer rewrite: "I was not included in the earlier discussion, so I may be missing context. Please send me the latest details and I'll help move it forward."

Same intent, far less damage. The point is rarely to be fake-polite; it's to say the firm thing without the line you would regret tomorrow.

Do It Privately: Drop Your Draft Here First

There's one problem with using a normal AI chat for this: your most emotional, half-finished drafts end up saved in a chat history you didn't really want. The draft you were ashamed of becomes an entry in your sidebar, sitting there until you remember to remove it.

Be precise about how long "until" is, because it varies. ChatGPT keeps ordinary chats until you delete them and then schedules deletion within about 30 days; its Temporary Chats stay out of your history but may still be retained for up to 30 days. Gemini's consumer activity defaults to 18-month retention, and conversations selected for human review can be kept for up to three years. None of that's permanent in the literal sense, and all of it's longer than the ten minutes you actually wanted the draft to exist for.

Secret Chat is built around exactly this habit — drop your draft here first, and it's private by default, so checking a message doesn't leave a trail:

Responses are deleted automatically where the provider supports it, so a tense draft isn't kept around after you're done with it.
Chats stay in your browser, stored locally rather than in a cloud account history.
Nothing is trained on your prompts — training is opted out under business API terms.
Your identity is anonymized — registration uses an email only, with no behavioral profile built from your messages.
You get a PDF privacy report per session, showing how the data was handled.

It also includes access to multiple AI models in one place, so you can compare how different models rewrite the same draft. To be clear about limits: it doesn't guarantee total privacy (the chosen model still processes your prompt to answer it), and it isn't a substitute for your lawyer, HR, or compliance team when a message has real legal weight. For the everyday "is this too harsh to send?" check, though, it's exactly the right tool.

If you also want to understand the ChatGPT side — Temporary Chat, turning off training, and what those settings really do — see how to use ChatGPT privately without saving chat history.

Better Than One Second Reader: Send the Draft to a Panel

Everything above assumes one second reader. That's already a large improvement on sending the first angry draft, but a single model gives you a single verdict, delivered with the same confidence whether the read is obvious or borderline. And borderline is exactly where this task lives: the study at the top of this article is about a gap humans themselves disagree about.

The AI Council is built for that. It sends one question (here, one draft) to several models from different companies at the same time, and reveals each answer the moment that model finishes, so you're reading the first read on your message while the rest are still writing. They're all asked at the same moment rather than one after another, so the check lasts about as long as its slowest model instead of the sum of the panel: the opposite of doing this by hand, where every wait begins only once the last one has ended. Then a cheap referee model pulls the factual claims out of all the answers and builds a disagreement table: which model asserts each claim, which one contradicts it, which never mentioned it at all. It adds a short synthesis of what at least two models agreed on, and a "verify before acting" list.

Applied to a tense message, that's a more useful object than any single rewrite. If three models independently flag the same sentence as the threatening one, you have your answer and you can stop deliberating. If one calls the draft firm and another calls it aggressive, the referee puts that disagreement in front of you by name, and that's the line worth rewriting, or the judgement worth making yourself rather than delegating. Silence is informative too: a claim only one model made, and the others never raised, is a read you should weigh rather than act on.

You don't have to hold four reads in your head to find that line, either, which is the other half of what a panel saves you. Four opinions on one draft are four screens of largely the same advice, with the one place they part company worded differently in each — exactly the thing you skim past when you're already annoyed and want the message gone. The referee returns a ready conclusion instead: a short synthesis of what at least two models independently agreed on, and a separate list of what stayed contested. What you read before you hit send is therefore shorter than a single model's rewrite, not four times longer, with the individual reads still there above it if a particular verdict looks wrong.

Three sizes are available. A Duet seats two models, a Trio three, and a Quartet the whole bench — ChatGPT, Claude, Gemini and Grok. On the Duet and Trio you pick which families sit at the table; the Quartet runs all four and has no picker. A Duet runs free within the daily free quota, on a lighter pair that answers from memory with no live web search, which is no loss for a tone check, since nothing about your draft is on the web. Credits buy the branded panels and the larger sizes. Where a question does turn on current or checkable facts, the whole panel answers with live web search attached, decided once for the panel rather than per model so the members stay comparable. You don't have to start over to use it, either: paste a rewrite you already have and let the panel check it, or switch an existing chat to a council and the panel reads the whole conversation, draft and all. Each answer's cost is shown next to the model that wrote it, with the referee on its own line.

Two honest limits, and they matter more on a work message than on a trivia question. Agreement between models is evidence, not proof — they share training data and can be wrong together, and the dialect and directness biases described further down this page are exactly the kind that several models can share; the reliable signal is disagreement, not consensus. And a council widens where your text travels: your draft reaches every model on the panel plus the referee. Each of those requests is anonymous (no profile, no account identity attached), but anonymization protects who you are, not what you type, so the "keep the details out of it" rule above applies with more force to a council, not less. The finished check is stored only on your device, and the Session Privacy Report lists every model that saw the draft.

Before You Hit Send

The whole habit comes down to one rule: when a message makes your jaw tighten, don't send it yet. Check it, understand why it reads the way it does, rewrite it calmly, and choose the version that protects both your point and the relationship. Do it in a private space so the messy drafts don't outlive the moment.

Before you send the message you might regret, drop your draft into Secret Chat first.

Where this workflow falls down

It would be easy to end here, so it's worth saying what the method doesn't do.

Tone detection is not solved. Sarcasm is where language models are weakest, and they degrade most on precisely the ambiguous cases where humans also disagree, which are the messages you most want help with. They also tend to read a message from the sender's point of view, the same egocentrism the study above describes.
It can misfire on how you write. Documented dialect and language-variety biases mean a non-native English speaker, or someone whose style is direct rather than hedged, may be told a normal sentence reads as rude. If the advice is to soften something you meant plainly, you're allowed to disagree with it.
Better tone is not the same as a better outcome. A 2026 randomised field experiment across six companies rewrote 16,880 real emails with an AI model. It shifted the measured positivity of the writing, and produced no change in open rates, reply rates or response time. (That's a preprint, so read it as a caution rather than a settled result.) The point of pausing is to avoid the message you would regret, not to engineer a reply.
Sending is close to final. Some tools offer undo-send or recall, but none of them reliably removes a message from notifications, someone's phone, an export or a compliance archive. Plan as though there's no recall.

Frequently Asked Questions

How do I check a message before sending it?
Write the draft but don't send it. Paste it into an AI with a prompt that asks how it comes across, which lines are risky, and for a calmer rewrite in two tones (polite and firm) then pick the version that keeps your point without the harsh edge.
Can AI rewrite my angry email?
Yes, and it's useful for exactly this, but treat it as a second reader rather than an oracle. Models are measurably weaker on sarcasm, particularly on the borderline cases humans themselves disagree about, and they tend to adopt the sender's perspective, which is the same bias you're trying to correct for. They also carry documented biases around dialects and language varieties, so a non-native English writer may be told a perfectly normal sentence sounds wrong. Remove names and identifying details first, and treat the output as a draft to review, not a final word on anything with legal weight.
What should I remove before pasting a work message?
Real names, company and client names, invoice numbers, addresses, and account details. The tone fix doesn't need them — describe the situation generically (for example, "a client refuses to pay an invoice"), and paste only the message itself.
Should I ask more than one AI to check the same draft?
It helps, and you don't have to do it by hand. Secret Chat's AI Council sends the draft to several models from different companies at once, shows each read as that model finishes, and then has a referee mark which model flagged each problem, which disagreed, and which never raised it, so a line three models independently call threatening is settled, and a line they split on is the one you decide yourself. A two-model Duet is free within the daily quota; credits buy the ChatGPT/Claude/Gemini/Grok panels and the larger Trio and Quartet. Two caveats: agreement is evidence, not proof, models share training data and can share the same tone and dialect biases, and a council sends your draft to every model on the panel plus the referee, so strip names and identifying details before you paste it.
Why use Secret Chat instead of a normal AI chat?
Because checking a tense draft in a normal chat leaves your most emotional drafts in a saved history. Secret Chat is private by default — local storage, training opted out, identity anonymized, automatic deletion requests, and a PDF privacy report — so the draft doesn't outlive the moment. The chosen model still processes your prompt to answer it.