You click “Help me write” in Gmail, type a one-liner like “decline politely, suggest next month,” and Gemini hands back a generic three-paragraph corporate email that opens with “I hope this message finds you well” and never actually references the thread you are replying to. Or the tone is off — too formal for a Slack-adjacent coworker, too breezy for a vendor. You polish manually, defeating the point.
The “Help me write” model behind the side panel is Gemini, but it runs with a much tighter context window than the standalone Gemini app, and its prompt template is template-driven rather than thread-aware unless certain conditions are met. This article walks through why the drafts come out off-tone and how to coax better output without leaving Gmail.
Common causes
Ordered by hit rate, highest first.
1. Gemini does not have access to the thread
The single biggest reason. “Help me write” inside a new compose window has no thread to read. Inside a reply, it should auto-attach the thread — but only if you opened the reply first, then clicked the Gemini icon. If you opened compose and then pasted the original email in, Gemini treats it as a blank draft.
How to spot it: drafts that ignore specific names, dates, or numbers from the thread are a tell. If Gemini writes “regarding our previous conversation” instead of “regarding the Q3 budget you sent Tuesday,” it does not see the thread.
2. Tone preset is set to “Formal” or unset
The “Refine” menu has Formalize / Elaborate / Shorten. The initial draft has no user-tone preset; it defaults to a neutral-formal voice tuned for B2B email. If your audience is a long-time coworker, this reads as cold.
How to spot it: opening line is “I hope this email finds you well” or “Thank you for reaching out.” That is the default tone signature.
3. Prompt is too short and too vague
“Decline politely” gives the model no constraints. It fills the gap with boilerplate. The “Help me write” UX rewards short prompts on the surface, but the underlying model needs the same scaffolding any LLM does: who you are to the recipient, what specifically you are declining, what you want to offer instead.
How to spot it: your prompt is under 10 words and the output is over 100. The ratio means the model is improvising 90% of the content.
4. Personal Google account, not Workspace
“Help me write” on a personal @gmail.com account uses a more conservative system prompt that biases toward neutral-formal English. Workspace accounts on Business Standard or higher get a slightly more flexible variant that picks up signals from prior sent emails.
How to spot it: same prompt on a Workspace account produces visibly different drafts. If you only have a personal account, you are on the more conservative path.
5. No prior sent-mail signal
Even on Workspace, the “match my style” pass needs a meaningful sample of your sent mail in the same account. A new Workspace seat with five sent emails has nothing to anchor to and falls back to default tone.
How to spot it: drafts feel generic regardless of prompt. Check the Sent folder — fewer than ~30 substantive emails means the style anchor is weak.
6. Wrong language detection
You write the prompt in English but the thread is in Chinese / Japanese / Spanish. The model may detect the prompt language and draft in English, or detect the thread and draft in the thread language — and it guesses wrong about which the user wanted.
How to spot it: output language does not match what you expected. Re-prompt explicitly: “reply in Mandarin” or “reply in English.”
7. Sensitive-topic redirect
If the thread mentions legal, medical, financial advice, or HR-coded keywords (termination, harassment, layoff), Gemini falls back to extremely cautious neutral-corporate phrasing regardless of your prompt. This is a safety pattern, not a tone bug.
How to spot it: drafts are noticeably stiffer for one specific thread but fine for others. Check the thread content for sensitive triggers.
Shortest path to fix
Step 1: Open the reply first, then click Help me write
Always click Reply on the thread, then click the pencil/Gemini icon inside the reply window. This attaches thread context. Do not click Compose and then try to paraphrase the thread into the prompt.
Step 2: Give a three-part prompt
Replace “decline politely” with three short clauses:
Decline the Tuesday meeting invite from Jane.
Reason: I am OOO that week.
Offer: any time the following Monday or Tuesday.Three lines, each one constraint. The model has enough to write a draft that references the thread and provides a concrete next step.
Step 3: Use Refine after the first draft, not before
Generate the draft, then use Refine → Shorten or Refine → Casual. The “Casual” refinement applies a tone overlay that the initial draft does not have. Two rounds of Refine usually lands the tone within one click of correct.
Step 4: Add a tone anchor in the prompt itself
Append “tone: like I am texting a coworker” or “tone: brief, friendly, no greeting line” to the prompt. The model honors explicit tone instructions more reliably than it infers from context.
Step 5: For personal accounts, paste a sample
If you do not have Workspace and the default tone is consistently wrong, paste two lines of one of your own past replies at the top of the prompt:
Match this voice: "yeah that works, lmk what time. I can do Tues after 3."
Now draft a reply that declines Wednesday and offers Thursday.This gives the model an in-prompt style anchor regardless of account type.
Step 6: Switch language explicitly
For mixed-language threads, state the output language: “Reply in English” or “用中文回复”. Do not rely on auto-detection.
Step 7: For sensitive threads, draft manually
If the thread touches legal / medical / HR topics, expect Gemini to refuse to drop the corporate tone. Either draft those replies by hand, or strip the sensitive keywords from your prompt and add them back in the final manual pass.
When this is not on you
Google ships A/B variants of the “Help me write” prompt template. Some weeks the default tone is noticeably warmer; other weeks it tightens up. If output quality regressed suddenly across all your threads with no change on your end, you are on a new variant. Wait a week or two and re-evaluate.
Workspace admins can also restrict Gemini features via the Admin Console. If “Help me write” is missing entirely, ask your admin.
Easy to misdiagnose as
Assuming the model is “just bad at email.” It is the same Gemini that drafts fine in the standalone app — the constraint is the side-panel template and context window, not model capability. Moving the same draft request to gemini.google.com with the thread pasted in usually produces a noticeably better result, which confirms the issue is panel-side, not model-side.
Also easy to confuse with “Smart Compose” — that is a different feature (the gray inline suggestions). Smart Compose is sentence-level autocomplete and uses a smaller, faster model.
Prevention
- Always open Reply first, then invoke Help me write — never the other way around.
- Use three-clause prompts: action, reason, offer.
- Append an explicit tone anchor for any draft going to someone you know well.
- For personal accounts, build a 2-line “voice sample” snippet you paste in for casual replies.
- For language-mixed threads, state the output language in the prompt.
- Accept that legal / HR / medical threads will draft stiffly — plan to hand-edit those.
FAQ
- Why does the standalone Gemini app write better email than the Gmail side panel? The app has a longer context window and a more flexible system prompt; the side panel is template-constrained for speed.
- Will Gemini learn my voice over time inside Gmail? Workspace accounts get some signal from sent mail, but it is not personalized fine-tuning — do not expect it to mimic you closely without an in-prompt anchor.
Related
- Gemini Extensions (Workspace, Maps, YouTube) Not Triggering
- Gemini Workspace Integration Issues
- Gemini Workspace Not Syncing
- Gemini Replies in the Wrong Language
- Gemini App and Web Give Different Answers
- Gemini Doc Summary Is Weak
- Gemini Google Account Permission Errors
- Gemini Gems Not Saving
- Gemini Saved Info Not Persisting
- Gemini Context Too Short
Tags: #Gemini #Troubleshooting #google-ai #Gmail #Workspace #help-me-write