Playbook, writing

Rewrite hard emails without losing your point

The blunt draft is your friend. Our routine keeps the point, fixes the tone, and gets the hard email out the door.

The Curio Team · 2 min readWritten with AI from the sources below and checked against them before publishing.

The hardest emails carry a no. Declining a project, chasing a late invoice, telling a client the date slipped. We write those backwards now: first a blunt draft that says exactly what we mean with zero softening, then a handoff where the model fixes the tone. The split is the trick. We own the message. The model owns the manners.¹

Tone control works when you name three things: who is reading, what relationship you are protecting, and what cannot change. Anthropic's docs call this giving the model a role, and OpenAI's prompting guide pushes the same idea from the other side: specific instructions with real context beat vague asks every single time.^{1 2}

Any current chat model handles this well, so use whichever tab is already open. The routine is the asset, not the brand. It's also the first move in learning to write with AI: separate what you say from how you say it, and never let a rewrite touch the what.

Here's the five-step routine we run on every hard email.

1. Write the blunt version first. Say the true thing in plain words: we can't take this on, the invoice is 60 days late, the launch moved. No softening yet. This draft is for you, not for them.
2. Give the model the scene. Paste the draft with one line of context: who the reader is, what you need them to do, and how much the relationship matters. A rewrite for your manager and a rewrite for a vendor are different emails.
3. Freeze what cannot change. Tell it the ask, the date, and the no are fixed. Everything else is negotiable. Without that rule, polite rewrites quietly drop the deadline.
4. Ask for two versions, one warm and one direct. Reading them side by side shows you the dial, and the best email is usually a sentence from each.
5. Read the winner out loud, then rewrite one sentence in your own words. A message that is all model sounds like a press release, and nobody believes a press release that says sorry.

The routine earns its keep on the worst messages. A refund refusal, a missed commitment, an apology that might get forwarded. For those we add one question: what could a hostile reader lift out of context? The model finds the screenshot sentence with unsettling reliability. We rewrite that sentence by hand.

We own the message. The model owns the manners.

One warning. Models over-soften by default, and a hard message wrapped in three paragraphs of appreciation reads as evasive, which lands worse than blunt ever would. The vendor guides push specificity for exactly this reason: vague prompts produce padded prose. If the rewrite buries your ask below the fold, cut the cushion and send the short one.²