Tone differences — warmer without losing intent

Same three models, very different tone defaults. The user wants a passive-aggressive deadline reminder rewritten warmer and more direct, under 80 words, no new commitments. Watch what each model thinks "warm" means.

The prompt

Rewrite this email warmer and more direct. Keep it under 80 words. Do not add new commitments.

Original: "Hi team, Per my last note, the deck is overdue. We need it by EOD or we cannot ship the campaign Monday. Please prioritise. Thanks."

Claude Sonnet 4.5

Hi team,

Just checking in on the deck we discussed. I know you're juggling a lot right now!

To keep our Monday campaign launch on track, we'll need it by end of day today.

Appreciate your help with this!

Thanks,

GPT-4o-mini

Subject: Quick Reminder: Deck Due Today

Hi team,

I hope you're all doing well! Just a friendly reminder that the deck is overdue. We really need it by the end of the day to ensure we can ship the campaign on Monday. Your help in prioritizing this is greatly appreciated!

Thanks so much!

Gemini 2.5 Flash

Hi team,

The deck is overdue

Side-by-side scorecard

What this tells us

Three completely different ideas of "warm". Claude opened with empathy ("I know you're juggling a lot right now"), softened the deadline language ("we'll need it"), and stripped the implicit threat. It also followed the constraint — no subject line, kept under 80 words, no new commitments added. Word count: 40.

GPT-4o-mini added a subject line the user did not ask for ("Quick Reminder: Deck Due Today"). Adding fields that were not requested is GPT-4o-mini's most consistent dialect quirk across this course — when it sees an email-shaped task, it volunteers email-shaped scaffolding. The body itself is warm but padded: "I hope you're all doing well" carries no information the recipient did not already have. Word count: 53 with the subject line, 47 without — either way inside the 80-word cap.

Note what the subject line does to the constraint. The cap was 80 words and GPT stayed under it, so nothing failed. But a rewrite task that returns an extra field is not a drop-in replacement for the original text: whatever code consumes this now has to decide whether to strip the first line. Unrequested scaffolding is a parsing problem before it is a style problem.

Gemini kept the greeting and then stopped: "Hi team," followed by a single un-rewritten clause lifted from the original. Six words. Nothing was made warmer, because nothing was rewritten.

Whether that is Gemini's behaviour or our capture's output cap is the open question from the previous lesson, and it is still open here — the stop reason was not recorded. The observable fact is narrower than "Flash returns un-transformed text under constraints": on this one run, the response ended after six words.

The takeaway for tone tasks: GPT and Claude both produce usable warmth, but they get there via opposite levers. Claude removes pressure — it deletes the threat and softens the deadline verb. GPT adds friendliness — it leaves the pressure in place and wraps it in pleasantries. Those are different edits, and which one your brand wants is a decision, not a quality ranking. If your voice is Bayt Coffee — direct, warm, no fluff — Claude's lever is the one you want. If your voice is American-corporate-friendly, GPT's is.

Captured 2026-04-27 from Claude Sonnet 4.5, GPT-4o-mini, and Gemini 2.5 Flash — stop reason not recorded. Re-runs may differ.

Next: how to choose the right model per task — a decision tree you can take to your team. :::