1 hour ago
5
Follow ZDNET: Add america arsenic a preferred source on Google.
ZDNET's cardinal takeaways
- Gemini powers Help Me Write for ratio successful Gmail.
- You tin usage Gemini oregon Claude to draught context-aware emails.
- Testing showed Claude is amended astatine matching code and requirements.
Three years ago, I started using Help Me Write, 1 of Google's archetypal attempts astatine applied AI. The penning adjunct is built into Google offerings similar Docs and Gmail, and I recovered it could draught full emails from scratch -- each I had to bash was motion disconnected connected them. As overmuch arsenic I similar Help Me Write, I've noticed the Gemini LLM underpinnings powering it someway present adjacent amended results done Google's standalone AI app.
Since I was feeding Gemini my email threads for drafting responses, I decided to comparison Claude's performance. After a week's worthy of testing, I amazingly recovered Anthropic's tools constitute messages a batch arsenic I would.
What I'm comparing, and however I'm investigating
These rival LLMs connection aggregate models, but I'm sticking to Gemini 3.5 Flash for speedy effect times, accessed done a Pro account. However, it's disposable for free, with wide usage limits. Importantly, I've switched connected Google Labs' Personal Intelligence toggle nether Attachments > More Tools > Personal Intelligence. It pulls actionable stylistic cues and accusation from my anterior Gemini chats, adding a idiosyncratic flair to the existent responses.
As for Anthropic's tool, I leveled the playing tract with the free Claude Sonnet 4.6 model, and Effort
set to High. It strikes a equilibrium betwixt velocity similar Gemini Flash, and all-important factual accuracy. Additionally, I dictate my prompts to prevention time, and Claude's speech-to-text motor is leagues up of Gemini, particularly if you person an accent.
Also: I fto Gemini Flows signifier my Gmail, and it efficaciously filtered my inbox (with 1 sneaky catch)
For each email, I outline the draft's objective, past travel with 2 modular instructions. They personalize results consistently and forestall token wastage. First, I inquire the AI to suspend contiguous responses and refrain from follow-up questions seeking details missing from the archetypal prompt. Second, I bid the LLM to imitate my code from the email thread. You'll request these to expect a workable response.
To illustrate, I'll usage an email pursuing up connected shipment delays and a missing tracking fig for a merchandise I program to review.
Here's the prompt:
"Draft a effect for the pursuing email thread and petition a shipment tracking fig erstwhile again, explaining that since I volition beryllium traveling successful mid-July, we request to enactment a timeline connected this merchandise reappraisal to guarantee it is done earlier I travel.
Do not respond immediately. Ask maine immoderate questions to draught a due response. Imitate my semiformal code seen successful the thread.
"<Email concatenation pasted/attached arsenic PDFs>""
Differences amusement immediately
Right disconnected the bat, Gemini asks maine 2 questions pursuing up connected the request. The archetypal seeks a legitimately utile item I didn't notation -- the targeted completion day for the review.
Sadly, Gemini's 2nd question asks if the recipient has replied since my past email successful the thread, but doesn't authorities wherefore it needs my answer. Questioning the supplied info seems redundant since I would've evidently included the latest response, and its accompanying justification remains ambiguous. The email thread includes a anterior petition for the tracking number, which the AI ignored.
"If helium did reply to your petition for the tracking number, knowing what helium said volition assistance maine code his latest points directly. If helium hasn't replied and you are pursuing up again, I volition framework it arsenic a gentle nudge. "
Also: Your Gmail is getting an AI makeover - here's what to expect and when
Claude performs similarly, but asks 3 questions instead. While Google's instrumentality asked for my planned reappraisal completion date, Anthropic requested the circumstantial question date, which is arguably much applicable and relatable accusation for my recipient. I cannot reasonably perpetrate to a completion day portion inactive awaiting the product, but the question day is simply a fixed constant.
The 2nd question highlights Claude's superior "understanding" of my prompts, straight asking maine however firm I privation to be. The justification explicitly states my effect volition power the email's tone, dissimilar Gemini's vagueness. Third, Claude pulled the websites I constitute for from my email's signature line, and asked maine wherever I would constitute the reappraisal successful question.
This item is irrelevant to the requested email draught and was referenced successful a anterior conversation, leaving maine mildly disappointed. To support investigating fair, I responded to each the questions from some AIs.
The email drafts are leagues isolated
Since this attack of including an full email thread with an LLM's punctual takes sizeable effort without integrations oregon scripted automations, I'd accidental my expectations were lofty. Currently, typing retired responses straight is conscionable faster than utilizing AI oregon improvising thing from Help Me Write successful Gmail, and I don't adjacent get that convenience successful non-Gmail accounts.
Gemini's effect covered each the bases, but ran 3 paragraphs long. They were structured well, with my tentative submission day successful bold. The code was formal, but friendlier than I'd similar utilizing successful the 2nd petition for the aforesaid info.
I've noticed that Gemini's responses are typically verbose, but you tin dodge that with Help Me Write, oregon power Personal Intelligence disconnected erstwhile you don't request to notation information from different chats. I recovered that the magnitude of this email draught diluted the urgency excessively much.
Gemini's response:
"Hope you're having a bully week.
I wanted to ellipse backmost connected this since I haven't heard from you regarding the tracking number. As mentioned before, a tracking ID oregon courier sanction would beryllium incredibly adjuvant truthful I tin support an oculus retired for customs clearance updates and guarantee the bundle reaches maine smoothly.
Additionally, I wanted to springiness you a heads-up regarding my timeline. I volition beryllium traveling retired of the state successful mid-July, truthful I americium targeting July 15th arsenic a hard deadline to wrapper up this merchandise review.
Because we are already adjacent to July, we are moving with a choky model erstwhile the illustration arrives. Could you delight cheque successful with Huanuo erstwhile much to unafraid that tracking info? Once I cognize precisely erstwhile it's landing, I tin officially fastener this into my contented calendar and warrant we get everything done earlier I travel.
Looking guardant to proceeding from you!
Best, Chandraveer"
Also: Anthropic rolls retired Claude Tag, your caller agentic AI coworker successful Slack
With Claude, the personalization and relatable cross-questioning laid the groundwork for a well-rounded draft. Like Gemini, it didn't omit facts, but it produced conscionable 2 paragraphs, person to the magnitude I would've written. Most importantly, the code and connection prime were splendid. While Google conscionable spat retired a draft, this AI highlighted the included pun "I'd privation the portion successful manus with capable runway to walk meaningful time" referring to my imminent travel.
A bundled taxable enactment and a Send via Gmail fastener were the cherry connected top. The second appeared since I logged into Claude with a Gmail ID, but it's thing you'd expect of Gemini instead, fixed its nonstop ecosystem integration.
Claude's response:
"Hope you're doing well! Just circling backmost connected the tracking fig - I inactive haven't received thing from a shipper connected my end, truthful it would beryllium large to person that item to enactment with.
I besides wanted to emblem thing worthy keeping successful mind: I'll beryllium traveling from the 15th of July, truthful I'd similar to person a reappraisal timeline locked successful good earlier that. Ideally, I'd privation the portion successful manus with capable runway to walk meaningful clip with it and constitute thing worthwhile. Could we hold connected a transportation deadline and reappraisal model that accounts for this? I don't privation america to beryllium scrambling adjacent to the wire.
Looking guardant to proceeding from you - and hopefully from a courier soon!
Best,
Chandraveer"
My caller preferred tool
It's nary concealed that Gemini benefits from nonstop integrations with Google's far-reaching ecosystem, but it dropped the ball, particularly fixed the prompt's wide correlation to emails. If you're utilizing a Gmail account, the Gemini LLM gets the occupation done with Help Me Write, but the results are verbose and vague. Using Gemini the roundabout mode isn't worthy the hassle.
In my tests, Claude routinely asked amended personalization questions to amended responses. Using it is simplified done integrations enabled successful the Skills section. So, if you're plugin-dependent anyway, and picking betwixt the two, I'd prime Claude.
English (US) ·