I Built an Email Assistant Gem — Here's What It Actually Does

Keywords: email assistant, AI email writing, Gemini Gem, professional email, vendor escalation, IT infrastructure, prompt engineering

Reading time: ~8 minutes

Last quarter I sent 247 work emails. I went back through my sent folder to count. Of those, about 30 messages took longer to write than they should have — escalation emails, vendor pushback, careful updates to non-technical stakeholders. The other 217 were quick. The 30 slow ones are what this post is about.

I run network operations for an ISP. Part of the job is sending messages where the wrong word choice creates real friction — a vendor interprets "concerned" as threatening, a client reads "minor delay" as "they don't know what they're doing," an internal escalation sounds like blame instead of fact. Most of these messages I rewrote three or four times before sending.

My Email Assistant Gem doesn't write those messages for me. It drafts them, I edit them, and the back-and-forth usually saves me about ten minutes per email. That sounds small. Multiplied by 30 messages a quarter, it's roughly five hours back in my week. Over a year, a full workday.

The Problem I Was Trying to Solve

Before I built this Gem, I had two ways to write a tricky email. I either drafted it from scratch and hoped the tone was right, or I pasted my draft into a generic chatbot with no context and got back something that sounded like a corporate apology.

The problem with generic AI email help is that it doesn't know my recipient. It doesn't know that the vendor I've been working with for three years has a different tolerance for directness than the new vendor I'm negotiating with for the first time. It doesn't know that my internal stakeholders include both engineers who want raw data and a director who wants a one-paragraph summary. Without that context, the AI drafts something safe — and "safe" usually means generic, slightly too formal, and reads like a template.

I needed something that already knew my context every time I asked.

What I Tried First (and Why It Wasn't Enough)

My first attempt was a long system prompt I pasted into Gemini's regular chat every time. Something like: "You are my email assistant. I work in IT infrastructure. My typical recipients are vendors, clients, and internal stakeholders. Draft professional email for me." It worked — sort of. But I had to paste the prompt every session, and Gemini would forget the context halfway through a long conversation. By the third draft in a single session, it was drifting back to default tone.

Then I tried ChatGPT's Custom GPTs. Same problem — better than pasting every time, but the context wasn't specific enough to my actual job. The drafts still needed significant editing.

What I wanted was an assistant that remembered my role, my recipients, my default tone, and my preferences — and could pull all of that in every new conversation without me re-pasting anything.

Gemini Gems did exactly that. The Gem holds the instructions permanently, so every conversation starts with the right context already loaded.

The Gem I Use

Here's the core of the instructions I saved into the Gem:

You are my email assistant. You draft and refine professional email for different audiences and purposes. When I describe an email I need to send:

  • Draft it in a tone appropriate to the relationship and objective I describe
  • Keep it concise — get to the point by the second sentence
  • Avoid filler phrases and corporate-speak
  • Offer one alternative version if the tone could reasonably go different ways
  • Flag if something I'm about to say could land poorly

My context: IT infrastructure professional. Regular recipients include vendors, clients, and internal stakeholders with varying technical backgrounds. Default tone: direct and professional, not formal.

The last paragraph is the part that matters most. It's what makes the Gem's responses calibrated to my work instead of generic email-speak. Without that paragraph, the drafts come back reading like they're written for someone in a different job.

Two Real Examples — Before and After

These are actual emails I sent in the last few months. I've removed company-specific details but kept the structure and tone changes.

Example 1: SLA Pushback to a Long-Term Vendor

Situation. A vendor we'd worked with for three years had missed their SLA response time three times in a month. The actual outages were minor, but the pattern was new. I needed to flag it without threatening to switch vendors — because we genuinely didn't want to switch, and they would read the email correctly.

What I asked the Gem:

"I need to tell a vendor that their SLA response time is unacceptable without ending the contract negotiation. We've worked together for three years and want to keep the relationship, but the pattern has changed. Draft that."

Gem's draft (edited version I actually sent):

Hi [Name],

I want to flag a pattern on SLA response times over the last month — three incidents, each outside the contracted window. None of the incidents were severe, but the frequency is new.

Can you look into what's changed operationally on your side? If there's a staffing or capacity issue, I'd rather know about it now than have it show up during a real incident. Happy to set up a quick call if that's easier than email.

— Justin

What the draft got right. It named the problem specifically (three incidents, outside the window) without listing the dates or quoting contract clauses, which would have read as adversarial. It offered a face-saving explanation ("if there's a staffing issue") instead of implying blame. And it ended with an open question, not a demand. The vendor responded within a day acknowledging the pattern and explaining a staffing gap. We didn't switch.

What I changed before sending. Roughly 10% — softened one phrase, removed the alternative version the Gem also drafted, added the vendor's name. The structure was the Gem's. The judgment about whether to send it was mine.

Example 2: Project Delay Update to a Non-Technical Client

Situation. A client's network upgrade was going to slip by two weeks due to a hardware shortage we couldn't control. The client had a board meeting where the original completion date was on the agenda. I needed to tell them early enough that they could adjust their own messaging, without making it sound like we didn't know what we were doing.

What I asked the Gem:

"Write a project status update for a non-technical client that explains we hit a delay without making it sound like we don't know what we're doing. Two-week slip, hardware shortage, not a planning failure. Original date was on their board agenda."

Gem's draft (edited version I actually sent):

Hi [Name],

Quick update on the upgrade timeline: we're looking at a two-week shift from the original [date]. The cause is a hardware allocation issue on the manufacturer side — not a planning or scope problem on our end.

The revised completion date is [date]. We've already adjusted our internal schedule so the slip doesn't compound.

If [date] is on a board agenda or external commitment, let me know and I can provide a one-line explanation that fits your messaging.

— Justin

What the draft got right. It separated the cause (manufacturer hardware) from us (planning) explicitly, which was the client's actual concern. It gave them a tool — a one-line explanation they could use internally. And it didn't apologize, which would have signaled more uncertainty than existed.

What I changed before sending. About 15% this time — added specific dates, removed an alternative version the Gem offered, and changed the closing line. The structure again was the Gem's.

Where This Works and Where It Doesn't

After six months of using this Gem for work email, here's an honest assessment.

It works well for:

  • Vendor and client emails where tone judgment matters more than facts. The Gem is good at calibrating tone once it knows the relationship.
  • Translating technical reality into language a non-technical stakeholder can act on. The alternative version feature is particularly useful here — I often end up sending the alternative instead of the first draft.
  • Reviewing a draft I've written myself. I paste my draft and ask "does this read as aggressive / apologetic / vague?" — the answer is usually accurate.
  • Status updates with a clear structure. The "name the problem, give the cause, state the next step" pattern works for most status updates.

It doesn't work well for:

  • Sensitive internal communications. Performance issues, HR-adjacent topics, anything that involves another person's reputation. The Gem can draft, but the tone judgment here should be human, not AI.
  • Emails that require specific historical context I haven't provided. If I'm referencing a conversation from six months ago, the Gem has no memory of it. I have to provide the context, which defeats some of the time savings.
  • Highly emotional situations. Apologies, conflict resolution, anything where the reader is already upset. The Gem defaults to a tone that's slightly too measured for these.
  • Emails where I don't actually know what I want to say yet. If I'm unclear on the message, the Gem will produce a polished version of unclear thinking. I have to know my point before I ask.

How to Set This Up Yourself

The setup takes about five minutes:

  1. Go to gemini.google.com/gems and create a new Gem.
  2. Paste the instructions from the previous section as the Gem's core instructions.
  3. Edit the "My context" paragraph to match your role, your typical recipients, and your default tone. This is the part that makes the drafts useful instead of generic.
  4. Test it with a real email you've been avoiding — that's the fastest way to see whether the tone calibration is right.
  5. Iterate. After your first five drafts, you'll know which instructions to adjust. The Gem's instructions are easy to edit.

The most common mistake I see people make with this kind of Gem is leaving the context paragraph too vague. "I'm a professional who writes email" produces generic drafts. "I'm a network operations lead for an ISP; my recipients are vendors I've worked with for years and new vendors I'm evaluating, plus internal engineers and external clients" produces drafts that need far less editing.

A Note on What This Replaces

This Gem doesn't replace the part of writing where you decide what you actually want to say. If I don't know whether to escalate or accommodate, no AI draft is going to make that decision for me. What it does replace is the mechanical work of turning a clear intention into a well-worded message — the part where you stare at a blank screen and rewrite the opening sentence four times.

For me, that's roughly five hours a quarter. The time savings aren't dramatic, but they're consistent — and for a task I'd rather not spend time on, that's enough.

Related Reading

Sources

— Justin

📅 First published: 2026-05-05 | 🔄 Last updated: 2026-06-21