When Claude can open Dropbox: build the review queue first

It is tempting to treat Dropbox's new Claude connection as another button: connect the files, ask AI, get the work done. For small teams, the more useful shift is simpler. When AI can find, summarize, save, and share files from the work folder, the team needs a clear review queue before anyone starts automating.

On June 25, Dropbox announced three new integrations for Claude: Dropbox Connector for Claude, Dropbox Plugin for Claude Cowork, and Dropbox Plugin for Claude Code. The idea is to let Claude work with files, feedback, and project material where they already live, with existing Dropbox permissions as the base layer.

Source: Work seamlessly with Dropbox in Claude, Dropbox Blog

For a small agency, shop, nonprofit, or school, this is not mainly a technical update. It is a work habit. If AI can save a summary back to Dropbox, create a sharing link, or organize a folder, you need clear labels: drafts, approved material, and files that must never leave the team without a human check.

What actually changes with Dropbox in Claude

Dropbox Connector for Claude is listed as a read-and-write connector. Claude can search Dropbox files, organize folders, save generated content, create shared links, and work with files the user already has access to. Anthropic's page also says the connector respects existing Dropbox permissions.

Source: Dropbox Connector for Claude, Claude by Anthropic

Dropbox Plugin for Claude goes further into everyday file workflows. The page describes search by name, keyword, type, or location, summaries of file content, saving Claude drafts back to Dropbox, file requests for uploads, and folder cleanup. Shared links require confirmation.

Source: Dropbox Plugin for Claude, Claude by Anthropic

It is easy to jump straight to the question: "What can we automate?" I would start one step earlier: "Which folder may AI help us with, and what does an approved delivery look like?"

An AI-connected review queue is a simple way of working where AI may help find, summarize, and suggest the next version, but every external delivery passes a clear human stop point. It sounds less exciting than an agent running on its own. It is also much easier to use on Monday morning.

Why this matters for small Nordic teams

Many small organizations do not have a document problem. They have a status mess. The right PDF is somewhere. The latest proposal is in a folder. The customer's comment is in meeting notes. The teacher's notes are in another file. Nobody is quite sure which version counts.

When Claude gets file access, AI can help with that boring middle work: search, compare, summarize, sort, and draft. But the same access can also make the mess faster. If AI saves new files with unclear names, creates shared links without a shared rule, or moves material before anyone reviews it, you do not become more efficient. You get a faster mess.

So today's practical task is not "connect everything". It is to build a small queue for one recurring workflow.

Choose work where files already shape the day:

• Proposals and customer material.
• Course material and student follow-up.
• Supplier invoices and receipts.
• Campaign material and feedback rounds.
• Policies, templates, and internal instructions.

That queue can be fully manual for the first week. The point is not that the tool does everything. The point is that the team can see the same status.

45 minutes: build the first review queue

Do this before you give AI broader file access.

1. Pick one recurring task. Write it as a sentence: "turn customer feedback into an action list", "summarize course evaluations", or "prepare proposal material".

2. Create a Dropbox folder for that task. Keep the name boring and clear, for example AI review - proposals - active.

3. Add four subfolders:

   • 01 Sources for material AI may read.
   • 02 AI drafts for summaries and suggestions.
   • 03 Human approved for material that may be used externally.
   • 99 Archive for old versions and material that should not steer the next answer.

4. Write a short README file in the main folder. It only needs to answer five questions:

   • What may AI help with here?
   • Which sources carry the most weight?
   • What must a human approve?
   • What must never be shared externally from this folder?
   • Who owns the workflow this week?

5. If you connect Dropbox to Claude, start with the bounded folder or a separate work account where possible. Review OAuth permissions, existing Dropbox access, and who may create shared links.

6. Run a test where AI may only suggest. No automatic sharing, no deletion, no moving original files. Save the draft in 02 AI drafts and let a human move it forward.

7. End with a simple run log: date, question, sources used, AI draft, human decision, and approver.

This is not paperwork for its own sake. It is what lets AI have real file access without making everyone nervous.

Copy-paste prompt: turn the Dropbox folder into a reviewed work queue

Use this prompt when Claude has access to the right Dropbox folder, or paste the same instruction if you are testing without a connector.

You will help us turn this Dropbox folder into a reviewed work queue.

Task: [write the task, for example summarize customer feedback before the weekly meeting]
Folder or sources you may use: [name the folder/sources]
You may do this: find relevant files, summarize, compare versions, and suggest next steps.
You may not do this without approval: share externally, create public links, delete files, move originals, or overwrite approved documents.

Do this:
1. List which sources you used and what each source appears to be.
2. Say if any source is old, contradictory, or missing.
3. Create a short draft with this structure:
   - Short status picture
   - Decisions needed
   - Questions for the human reviewer
   - Suggested next files or links to create
4. Mark everything that needs human approval before it is sent, shared, or saved in the approved folder.
5. End with a run log: date, sources, uncertainties, and recommended approver.

The important line is "suggest, but do not act without approval". Once that habit is in place, you can later let AI do more: save a draft, create a file request, or prepare a sharing link. But everyone knows where the stop point is.

Three small workflows to test

Proposal without version chaos

Put the latest customer email, old proposal template, and pricing material in 01 Sources. Ask Claude to find what changed, suggest questions for the customer, and draft the proposal. A human checks price, terms, and promises before anything lands in 03 Human approved.

School follow-up without hunting for notes

Collect the week's student or course notes in a bounded folder. Ask AI to sort recurring blockers, suggest the next support step, and mark what requires the teacher's judgment. Do not publish anything automatically. The point is to save teacher time before the meeting, not to let AI make pedagogical decisions.

Customer support with clear sources

Put the policy, price list, and previously approved answers in the source folder. Ask AI to write a reply draft with source references and a list of items that need a human decision. If the answer later connects to a ticketing system, use scoped API keys, keep secrets in environment variables or a secret manager, and route every external answer through approval.

Safer integration without slowing everything down

You can integrate AI without making the whole project defensive. Start with small, visible permissions.

• Give read access when that is enough.
• Share one project folder, not the whole Dropbox account.
• Use separate service accounts or scoped API keys for workflows that connect onward to CRM, calendar, or ticketing systems.
• Keep keys in environment variables or a secret manager, not in chat.
• Redact unnecessary personal data in test workflows.
• Require approval before AI sends, shares, deletes, moves originals, or creates external links.
• Keep a short run log so you can see what AI used and what the human decided.

This is a good first Tool Forge project: not a grand transformation, just one concrete file workflow where AI helps without owning the decision. Once the queue works, Skill Forge can make the routine repeatable: templates, prompts, naming, permissions, and follow-up.

What should be true after one week

After one week, you should be able to answer five questions without searching:

• Which folder may AI use for this workflow?
• Which files are sources, which are AI drafts, and which are human approved?
• Who may create or approve a sharing link?
• Which decisions may AI never take on its own?
• Where is the run log?

If you can answer that, you have gone further than many teams that "test AI" in ten different chat windows. You have built a place where the work can actually continue.