Save recurring email attachments to Drive with Claude

“The same report lands in my inbox every week and I file it by hand, or I forget to.”

Attachments filed into the right Drive folder, without the files ever passing through the conversation.

Tools it uses

gmail_searchgmail_readgmail_save_attachment_to_drivedrive_create_folder

What it needs

Google Workspace connected
Works with one connected account
The gateway URL in your Claude client

This used to be an Apps Script. Someone wrote it, it worked, and then it quietly broke and nobody noticed for a month.

The interesting part is what does not happen: the attachment never enters the conversation. It goes from Gmail to Drive server-side, so a 5MB spreadsheet costs you nothing in context.

SKILL.md

---
name: attachments-to-drive
description: File recurring email attachments into Drive folders
---

# Attachments to Drive

## Find the mail, not the attachment

Search narrowly. `gmail_search` with the sender and a file condition:

    from:[email protected] has:attachment newer_than:14d

Prefer `from:` and `subject:` over body text. Recurring reports are recurring
precisely because they come from the same place with the same subject.

## Read only if you need to decide

`gmail_read` with `text_only: true` when you need the body to route the file —
a period, an invoice number, which region it covers.

DO NOT read a message without `text_only` when it has attachments. The raw
MIME payload includes the attachment base64-encoded, and base64 is about a
third larger than the file. A 5MB CSV arrives as roughly 6.7MB of text, which
is a whole context window spent on a file you were never going to read.

The compact view still lists each attachment's filename, mimeType and
attachmentId, which is everything you need to file it.

## File it

`drive_create_folder` first if the destination does not exist. Then
`gmail_save_attachment_to_drive` with the message id, the attachment id and
the folder id.

The file moves Gmail to Drive server-side. Nothing about its contents reaches
this conversation, which is the point — you are filing it, not reading it.

## Rules that keep this from going wrong

1. **One folder per cadence, not per file.** `2026/Q3/vendor-reports`, not a
   folder per report. You are building somewhere to look, not an archive of
   directories.
2. **Check before you create.** Creating a folder that already exists gives
   you two folders with the same name and no error, and Drive will let you
   keep doing it.
3. **Never search the whole mailbox.** `has:attachment` alone matches years of
   mail. Always bound it with a sender and a date.
4. **File, then confirm, then mark.** If you also mark the mail read, do it
   after the file lands, not before — otherwise a failure leaves you with a
   read message and no file, which is invisible.

Paste this into your Claude client as a skill, then ask for it by name. It calls your connected accounts through the gateway, and every write asks you first.

Notes from running this

The base64 arithmetic is the whole reason this tool exists. Attachments are base64-encoded in the MIME payload, which inflates them by roughly a third. Reading one 5MB attachment the naive way costs about 6.7MB of context. The save-to-Drive path never puts the bytes in the conversation at all, so the cost is the same whether the file is 5KB or 50MB.

Responses are truncated before they can overflow. The connector caps tool responses around 900KB because MCP starts failing near 1MB. That is a backstop, not a plan — a truncated response is a response you cannot trust, so the fix is to not ask for the payload in the first place.

text_only is not just smaller, it is differently shaped. It flattens the headers and gives you attachment metadata as a list. Getting the same information out of the raw payload means walking a nested MIME tree.

Frequently asked questions

Does the attachment pass through the conversation? #

No, and that is the point of this skill. The file moves from Gmail to Drive server-side, so a 5MB spreadsheet costs nothing in context and the cost is the same whether the file is 5KB or 50MB.

Why not just read the attachment and save it? #

Because of the arithmetic. Attachments are base64 encoded in the MIME payload, which inflates them by roughly a third, so reading one 5MB attachment the naive way spends about 6.7MB of context on a file nobody was going to read.

Will it create duplicate folders in Drive? #

Not if it checks first, which is why the skill says to. Creating a folder that already exists gives you two folders with the same name and no error at all, and Drive will let you keep doing it, so the pattern is one folder per cadence rather than one per file.

Will it search my whole mailbox? #

No. Searching for attachments alone matches years of mail, so every search here is bounded by a sender and a date.

What order does it do things in? #

File, then confirm, then mark. If the mail is also marked read, that happens after the file has landed, because a failure the other way round leaves a read message and no file, which is invisible.

Run it against your own data

The agent runs this skill for you, on the accounts you connect, and shows you what it did.

Run this skill

By continuing you agree to our Terms of Service and Privacy Policy.

Other skills

Get a morning brief across your mail, calendar and tasks with Claude

One self-addressed brief: today's events across every calendar with the conflicts called out, the mail that needs a human grouped by account, and a task created for each thing you owe someone, with the noise labelled and marked read so it can be audited or undone.

Triage your Gmail inbox with Claude

A clean inbox and one self-addressed digest of what actually needs you, labelled so every action can be audited or undone.