| name | google-drive-organization |
| description | Survey a Google Drive's actual folder structure, apply a PARA-lite organization framework, and produce (then safely execute) a concrete reorganization plan โ deduping, sweeping root-level clutter, and fixing misnested folders. |
Google Drive Organization
Reorganizing a personal Drive well means surveying what's actually there before
proposing structure โ a generic "PARA folder template" imposed on an unsurveyed Drive
just creates new clutter next to the old clutter. This skill is a two-phase process:
survey + plan, then (separately, with explicit go-ahead) execute.
When to use this
- The user asks to organize, clean up, or restructure their Google Drive
- A Drive has grown organically for years and has root-level file dumping, duplicate
folders, or inconsistent naming
- Before a big life change that generates a new document category (new job, new
property, new year) and the user wants to know where things should go
Phase 1 โ Survey (read-only, delegate to an agent)
Use the Drive MCP tools (search_files, list_recent_files, get_file_metadata) โ
strictly read-only, no creates/moves/deletes during survey. A single research/general
subagent can do this well since it's read-only and the raw survey output (dozens of
folder listings) doesn't need to stay in the main conversation's context.
What to actually look for โ these are the recurring real problems, not hypothetical
ones:
- Root-level flat dumping: files sitting directly in "My Drive" root with no
folder. Query
parentId = 'root' and separate files from folders.
- Duplicate files: same or near-identical
fileSize under similar titles, often
created minutes apart (a sign of an accidental double-upload). Compare fileSize
and createdTime, don't assume โ verify by comparing at least the first page of
contentSnippet too if titles differ slightly.
- Misnested folders: a year-folder or category folder living inside a sibling
instead of next to it. Confirm actual nesting before reporting it as a finding โ don't
trust folder names alone, and don't trust a single opaque-ID lookup either (see the
verification note below). A real example that turned out to be a false positive: an
agent claimed a
2024 Taxes folder was nested inside 2025 Taxes based on a
parentId match; re-checking with rclone lsf "remote:Taxes" --max-depth 2 showed
2024/2025/2026 Taxes were all correctly siblings โ the ID had been misattributed to
the wrong parent folder during verification.
- Fragmented topic folders: many single-purpose top-level folders that are really
one topic (e.g., 7 different
Health Sync <metric> folders instead of one Health/
folder with subfolders).
- Overlapping/duplicate-purpose folders: a top-level folder and a subfolder
elsewhere covering the same subject (e.g., a top-level
Remodel/ next to
<Property>/Kitchen Remodel/).
- Naming drift: trailing spaces, inconsistent casing, ambiguous dates
(MM-DD-YYYY vs YYYY-MM-DD) โ cheap to fix, good signal of no naming convention.
- Stale/inactive folders: old hobby folders, completed projects (e.g., a home
search folder after the purchase closed) still sitting at the same top level as
active work โ PARA's "Archive" category exists exactly for these.
- Unmanaged
Shared with me: items never triaged into the real folder structure.
Google Drive has no native way to fold these into folders โ the only mechanism is
"Add shortcut to Drive" for the ones worth keeping.
Verify surprising/specific findings (a claimed duplicate, a claimed misnesting) with a
quick follow-up search_files or get_file_metadata call yourself before presenting
them as fact โ a subagent survey can still get specific file IDs or paths wrong.
Phase 2 โ Framework
PARA (Projects / Areas / Resources / Archive) is the right default for a personal
life-admin Drive โ it organizes by actionability rather than topic:
- Projects: time-boxed efforts with a goal and an end (a home renovation, a house
search, a wedding)
- Areas: ongoing responsibilities with no end date (finances, health, a property
you own)
- Resources: reference material that isn't currently actionable (manuals, books,
research)
- Archive: anything from the first three that's gone inactive โ moved out, not
deleted
The critical habit most unorganized Drives are missing isn't folder creation, it's
archiving on completion โ when a project (house search, a specific renovation
phase) finishes, its folder should move to Archive, not sit forever at the same top
level as active work.
Don't reach for a numeric system like Johnny Decimal unless the Drive has dozens of
top-level categories that PARA's four buckets can't reasonably absorb โ for a single
person's life-admin Drive, a hard cap on top-level folder count (aim for well under
10) usually gets you the "nothing is more than a couple clicks away" benefit without
the addressing overhead.
Naming conventions (apply going forward, don't retroactively rewrite years of
existing filenames for marginal benefit):
- Date-bound files/folders:
YYYY-MM-DD prefix โ the only format that sorts
correctly both chronologically and alphabetically across every OS/locale
- Year-based categories (tax folders, annual exports) as siblings, never nested
inside each other
- Status tags in brackets (
[FINAL], [DRAFT]) when a file's state matters and isn't
obvious from content
- Prefer hyphens/underscores over spaces for portability; watch for trailing spaces
from manual folder creation (a real, common source of near-duplicate-looking folders)
Phase 3 โ Produce the plan
Deliver: a proposed top-level taxonomy (aim for ~5-8 categories mapping to the
person's actual areas of life), a migration table (current path โ proposed path) that
maps every existing top-level folder somewhere, and a short list of dedup/misnesting
fixes to make before any bulk moves. Be honest about scope โ if the survey shows a
Drive that's already reasonably organized, say so and propose a light-touch cleanup
rather than inventing a full rebuild to justify the exercise.
Present the plan for review (an Artifact works well for a long migration table) before
executing anything.
Phase 4 โ Execute (only with explicit go-ahead, in stages)
Moving and deleting real files/folders needs the user's explicit permission per
session safety norms โ this phase is not something to run unattended even if the plan
itself was approved in principle:
- Dedup fixes first โ these are the highest-value, lowest-risk items (confirmed
duplicate files/folders). Still confirm before deleting anything; there's often no
"delete" capability exposed via MCP tools at all, only via the Drive web UI, which
moves to Trash (reversible) rather than permanently deleting.
- Misnesting fixes โ move a folder to its correct parent.
- Root sweep โ move loose root-level files into their target folders, in batches
the user can sanity-check (e.g., by category), not all at once.
- Rename passes โ trailing spaces, date-prefix conventions going forward.
- Re-survey afterward to confirm the moves landed where intended โ don't assume a
move succeeded without checking, the same way an upload shouldn't be assumed to
have succeeded without checking (see the file-integrity lesson below).
Prefer rclone over the Drive MCP tools when it's available
Check rclone listremotes first โ if a gdrive:-style remote is configured (verify
it's the right account by comparing rclone lsf gdrive: --max-depth 1 against known
folder names, since rclone about's quota won't match the Drive web UI's combined
Google One figure), it beats the MCP tools for most of this work:
- Uploads:
rclone copy <local file> "gdrive:path/to/folder/" is a real,
purpose-built multipart upload โ no base64 encoding, no size limits to silently hit.
- Verification:
rclone lsf "gdrive:Some Folder" --max-depth 2 and rclone lsl
(with sizes) are simple, human-readable path-based listings โ far easier to sanity
check by eye than chasing opaque parentId values through the MCP search_files
tool, where it's easy to misattribute which ID belongs to which folder (see the
misnesting false-positive above โ that mistake came directly from this).
- Still verify byte sizes after an
rclone copy the same way you would for any upload
โ don't skip verification just because the tool is more trustworthy.
A hard-won lesson: verify uploads/moves, don't trust agent reports
If rclone isn't available and uploading falls back to the create_file MCP tool with
base64 content: that tool has been observed to silently produce corrupted/truncated
files when the upload work is delegated to an agent that spawns its own sub-agents to
parallelize it โ files arrived at ~1500 bytes regardless of their real size, with no
error thrown. If delegating uploads this way:
- Explicitly forbid the agent from using the Agent/Task tool itself โ require it to
do every upload sequentially, itself
- Require a size-verification step after every single file (
get_file_metadata's
fileSize compared against the local file's real byte count via stat), not just
a final "did files appear" check โ a corrupted file still shows up in a folder
listing
- Independently re-verify the agent's final report yourself with your own
search_files/get_file_metadata call before telling the user anything is done
This whole section is why the rclone path above is strongly preferred when available.
References