I never thought I'd be filming content in my car. Very heavy influencer vibes. The funnier part: I don't even have a commute.
Interactive voice AI isn't really my thing either. I'm a one-way dictation person — after a workshop I talk my thoughts and follow-ups into my phone, and my AI setup picks them up from there. No back-and-forth needed.
But several clients I've worked with love Claude's interactive voice mode, and they have real commutes. The same pattern kept coming up in our conversations: dead time in the car, and twenty, thirty emails waiting at the desk afterwards. This tutorial is for them — and for anyone who wants to handle email while their hands are busy, whether that's the car, the gym, or the supermarket queue. So I built the workflow and filmed it where it lives: in the car.
Here's what it does. Claude reads your inbox to you one email at a time, you decide out loud what happens to each one, and by the time you park the inbox is sorted: replies drafted, tasks captured, information filed. Nothing gets sent without you. That was the plan, anyway — more on the one email that slipped through below.
This works because Claude's voice mode can use connectors — the official links between Claude and apps like Gmail — while you're talking to it. At the time of filming, it's the only voice assistant I know of that can. The whole setup is one Google Doc and three connections. No code. The video walks through it start to finish, including the live run from the driver's seat, hiccup included.
The four lanes
Every email in your inbox needs exactly one of four things from you:
- Reply — you dictate the gist, Claude writes it properly and saves it as a Gmail draft.
- File — the key information gets saved as a doc in your Drive.
- Task — it lands on your to-do list, with a due date if the email implies one.
- Forward — Claude sends it on with a one-line note. This is the lane that bit me on camera. Keep reading.
(And "archive it" — the fifth, unofficial lane everyone recognizes.)
The triage session is deciding lanes out loud, not doing desk work by voice. The point is that most of the desk work has already happened by the time you sit down.
Step 1: write the instructions once
You could explain your rules to Claude fresh every morning, or write them down once. I wrote them down.
Ideally this would be a Claude skill — a reusable instruction file Claude loads when a particular job comes up. At the time of filming, voice mode can't use skills. So I did the next best thing: wrote the instructions like a skill and saved them as a Google Doc in my Drive. Claude can read Drive through a connector, so in the car I can say "there are instructions for you in my Google Drive" and it finds the doc, reads it, and follows it. When skills do arrive in voice mode, the doc moves over as-is.
I didn't write the doc by hand. I told Claude what it should contain and how I wanted it to work, and let Claude write and format it. It knows how a skill should look better than I do.
What's in it: a name and a one-liner about when to use it, the session flow (unread inbox, one email at a time, oldest first, a short brief per email, say "next" and move on), the four lanes, a couple of guidelines (draft, never send; ask when unsure), and a wrap-up where Claude checks its own work and summarises what it did. That's the whole doc.
Here's mine, with the two lines you'd change marked in [BRACKETS]. Copy it into a Google Doc named "Commute triage instructions" — that exact name is how Claude finds it by voice. (It also lives on its own page: The voice email triage instructions.)
**Name:** Commute email triage
**Description:** Use these instructions whenever I ask you to triage my inbox by voice. I'm usually commuting and may be driving, so keep everything short and audio-friendly.
## Session flow
- Go through my unread inbox one email at a time, oldest first.
- For each email: tell me who it's from and give me a one-or-two-sentence summary. Then wait for my decision.
- If I ask you to "read it in full", read the whole email.
- After each email is handled, say "next" and move on. No recaps between emails.
## My four lanes — every email gets exactly one
1. **Reply** — I'll dictate the gist; you write it properly in my tone. Read the draft back to me once. Save it as a draft in Gmail. Never send anything.
2. **File** — save the key information as a doc in Google Drive. Default folder: [YOUR DRIVE FOLDER] — I'll name a different folder when it matters.
3. **Task** — add it to Todoist (swap in your own task app's name here if you use a different one) with a due date if the email implies one. Default project: [YOUR TASK PROJECT].
4. **Forward** — prepare a forward with a one-line note from me. Save it as a draft. Never send.
"Archive it" is also a valid decision — archive and move on without asking follow-ups.
## Guidelines
- Draft, never send. No reply, forward, or anything else leaves my account during this session. I review and send at my desk.
- If an email doesn't fit any lane, or you're unsure, say so and leave it unread — I'll handle it at my desk.
- Don't invent details in replies. If you'd need information I haven't given you, ask me in one short question.
## Wrap-up
When the inbox is done, give me a short summary: how many emails went into each lane, and anything left unread for my desk.
The two decisions worth thinking hard about: which lanes you want, and which tools Claude needs access to for each one. Everything else in the doc is plumbing.
Step 2: connect the tools, then decide what they're allowed to do
Three connectors:
- Gmail — reads the inbox, writes the drafts.
- Google Drive — reads the instructions doc, and it's where the file lane puts things.
- Todoist — the task lane. I use Todoist; if your to-do list lives somewhere else, browse the connector directory. There's a good chance it's there.
In the Claude desktop app, go to Customize > Connectors. The Connected tab shows what you already have; Browse is where you add new ones. Each connection is a normal sign-in screen. I set everything up on the desktop because it's easier to click through — the connectors carry over to the phone.
The part that matters comes after the sign-in. Click into a connector and you get a list of every tool it gives Claude, split into read-only tools and write/delete tools. Each tool has three settings: always allow (Claude does it without asking), needs approval (you get a pop-up first), or blocked (Claude can't even ask). Here's how I set Gmail:
- Reading tools: always allow. That's the default, and it's fine.
- Create a draft: always allow. Otherwise you're approving a pop-up while driving.
- Reply to an email: needs approval.
- Send an email: blocked.
- Add a trash or spam label: needs approval.
Hold on to that distinction, because the rest of this tutorial turns on it: these settings are the rules. The instructions doc is a request. It took me one forwarded invoice to feel the difference.
Step 3: run it
In the car — parked, in my case, you can suspend your disbelief — open Claude, start voice mode, and say something like: "Help me triage my inbox. There are instructions for you in my Google Drive." Claude finds the doc, then pulls the unread inbox. Mine had 17 unread threads going back to early August. The oldest was a newsletter, and Claude offered to skip straight to the four from today that looked time-sensitive. Yes please.
From there it's a rhythm. Decision, "next", decision, "next":
- Mette asks to push Thursday's call to the afternoon. Reply: "Thursday works, suggest 2 p.m." Claude writes it properly and saves it to drafts.
- Jonas in accounting wants the Q3 receipts by next Friday. Task: into Todoist, Work admin project, due Friday.
- Signe confirms the workshop venue and needs a final headcount by the 8th. Both: file the confirmation to my Workshops folder in Drive, and a headcount reminder on tomorrow's list.
- An invoice that belongs with our invoicing address. Forward. This is where it went sideways — next section.
One practical thing from that first reply: the first time Claude wants to create a draft, you get a permission pop-up. Run the flow once at your desk and click "always allow" there, so it never happens while you're driving.
Back at the desk, the proof: the reply sitting in Gmail drafts, two tasks in Todoist with the right dates, and a venue confirmation doc in the Workshops folder. Everything the instructions promised, plus one thing they didn't.
Epilogue: guidelines vs rules
For the invoice, I asked Claude to forward it. Claude asked for permission to use the forward tool. I'd never seen it draft a forward, I was curious whether it would draft or send, so I approved it once. It sent.
Claude noticed before I did: "I need to flag something important. That forward actually sent immediately rather than saving as a draft. Your instructions say never to send anything, only draft, and I should have verified the tool's behavior before using it."
No harm done — the address was a made-up one inside my own domain. And I kept the whole thing in the video, because it's the best lesson in it. The instructions said "draft, never send" in plain language, and Claude followed them email after email. Then a tool that only knows how to send got approved, and the instructions lost. Partly the tool's limit, partly mine: I approved something without fully understanding what I was giving it access to.
Prompts for AI are guidelines. They are not hard and fast rules. The rules are the external limits you put on the system — in this case, which tools Claude is allowed to use. I had given it permission to use the forward tool. I could have blocked it, and that is a far more robust way of stopping Claude from sending email than writing "never send" in a doc.
So that's the change I'd make before your first commute: block the tools that send, forwarding included. The forward lane becomes "label it, I'll forward it from my desk". Where you draw the line is up to you. The permission settings are where it holds.
Make it yours
If your stack looks different — Outlook instead of Gmail, another task app, a train instead of a car — paste the AI onboarding doc into your favorite AI assistant. It interviews you about your email, your docs, and your task app, then builds your own version of the instructions, permission settings included.
- The voice email triage instructions — my doc, copy-paste ready.
- Build your own voice email triage assistant — the AI onboarding doc.