Stop uploading files to AI. Build a source folder first

The new small annoyance in AI work is not that the model is too weak. It is that we feed it the wrong pile of files.

A proposal sits in Dropbox. An old customer thread is in Gmail. A meeting note lives in a document nobody remembers by name. Then someone opens ChatGPT, uploads three files by guesswork, and asks for a good answer. Sometimes it works. Often it creates a polished draft based on half the truth.

Dropbox has announced three new apps in ChatGPT: Dropbox, Dropbox Dash, and Reclaim AI. The Dropbox app is meant to let users access, preview, save, and share Dropbox files from ChatGPT. The Dash app brings knowledge from more than 30 work apps into the conversation. The Reclaim app helps with calendars, meeting times, and scheduling conflicts. Dropbox also says existing sharing permissions and access controls continue to apply.

Source: Move your work forward with new Dropbox apps in ChatGPT

For a small Nordic team, this is not mainly an app announcement. It is a useful reminder: before AI gets access to work material, the team has to decide what counts as the right sources.

Today's AI input: the sources move into the chat

When AI can read from storage, search across work apps, and put the result back into the same systems, some manual copy-paste work disappears. Good. But sloppy context becomes more expensive.

If a salon, small accounting firm, association, or school connects documents to AI without a simple structure, AI may find fresh versions, old versions, internal notes, half-finished templates, and files that should not guide the answer. The issue is not that AI gets context. It is that the context is not curated.

A source folder is a bounded set of documents that the team actually wants AI to use for a specific piece of work. It might be a Dropbox folder, a Dash Stack, a Notion page, a Drive folder, or just a clearly named project folder. The format matters less than the rule: these are the sources that count for this question.

Dropbox Dash describes Stacks as collections where teams can group project briefs, links, messages, and meetings, share the collection, and use AI to summarize or ask questions about that content. That is a good pattern even for teams without Dash: collect first, ask second.

Source: Dropbox Dash Stacks

Learn this: do not connect the whole mess first

There is a tempting but weak sequence:

1. connect all the files
2. ask a large question
3. hope AI finds the right material

A better sequence is less exciting and much more useful:

1. choose one real workflow
2. create a source folder for that workflow
3. remove old versions
4. write a simple usage rule
5. let AI produce a first draft
6. review sources, decisions, and next steps before anything moves on

Dropbox Dash Chat emphasizes answers grounded in connected files and tools, with sources that help the user check the answer. Small teams should use that as the bar. An AI answer without sources can still be useful for ideas. An AI answer that affects a customer, student, contract, proposal, or schedule needs to show what it is based on.

Source: Dropbox Dash Chat

This matters even more when AI can do more than write text. If it can save, share, or schedule, a good prompt is not enough. You need a small operating routine.

45-minute test: build a source folder before the next AI answer

Choose one recurring case. Not everything. One.

Good candidates:

• A proposal request
• A new course registration
• A recurring support email
• A school question that needs a summary
• A customer follow-up after a meeting

Then do this for 45 minutes.

Minute 0-10: choose the case and create the folder

Create a folder with a name like Sources - proposal questions - active. Add only material that is allowed to guide the AI answer. If old versions exist, move them into a subfolder called Archive - do not use without checking.

Minute 10-20: write a README at the top

The README does not need to be elegant. It should answer four questions:

• What may AI use this folder for?
• Which documents matter most?
• What must a human approve?
• What must never be sent directly to a customer, student, or supplier?

Minute 20-30: mark owner and date

Add an owner to the folder or filenames where needed. Write when the sources were last reviewed. If a price list, policy, or course plan has no date, AI can easily use the wrong version.

Minute 30-40: run one AI test

Ask AI to answer a real question using only this source folder. Force it to show sources, uncertainties, and a short list of things that require manual checking.

Minute 40-45: decide whether the workflow deserves integration

If the test saves time and stays accurate, document the routine. If it needs too much cleanup, improve the source folder before connecting more apps.

Copy-paste prompt: ask the folder before you ask the world

Copy the prompt and replace the brackets.

You are helping me use a bounded source folder for a real work case.

Work case:
[Describe the case: proposal request, support email, course registration, customer follow-up, etc.]

Allowed sources:
[List the folder, stack, or documents that may be used]

Task:
1. Answer only from the allowed sources.
2. Write a short draft that I can review.
3. Show which sources you used under each important conclusion.
4. List anything missing, unclear, or requiring a human decision.
5. Suggest the next step, but do not send anything or change anything without my approval.

Format:
- Short summary
- Draft reply or work order
- Sources used
- Uncertainties
- Decisions a human must make
- Next step

It is not a magic prompt. The point is that it does two things many teams skip: it limits the sources and makes uncertainty visible.

Three small workflows worth testing

The proposal folder

Collect the latest price list, delivery terms, three approved past proposals, and a short note about what you do not promise. Ask AI for a first proposal draft with assumptions marked. A human approves price, scope, and wording before anything is sent.

The school folder

Collect the course plan, lesson outline, previous parent or student information letters, and a checklist for adaptations. Ask AI for a first information draft or lesson summary. The teacher checks tone, level, and what may actually be communicated.

The customer support folder

Collect FAQs, return terms, troubleshooting steps, and examples of good replies. Ask AI to suggest a reply and an internal note. A person reviews before the answer goes out, especially if money, contracts, or an unhappy customer is involved.

Integrate without opening the whole storage room

The safe way to start is not to avoid integration. It is to integrate narrowly.

Use read access where that is enough. Share a project folder, not the whole storage account. If you build API workflows around Dropbox, calendars, or CRM: use scoped permissions, separate API keys, environment variables, or a secret manager. Add redaction of personal data where possible. Require human approval before AI can send, save, or share externally. Keep a simple run log: question, sources, draft, decision, who approved.

That may sound bureaucratic until something goes wrong. Then the log is the difference between "AI did it" and "we know exactly which source and which decision led here".

What Hammer would build around this

In Mindset Forge, we start with the rule: which sources may AI trust, and when must a human decide?

In Tool Forge, we build the workflow: the right folder, permissions, prompt, and connection to Dropbox, Dash, ChatGPT, calendar, or CRM.

In Skill Forge, we make the routine repeatable for the team. Not as a long policy. As a short work habit: create source folder, ask with source requirements, review uncertainties, approve, log.

That is less shiny than a big AI agent. It is also much closer to daily work. For small teams, the next AI step is often not a larger model. It is a better pile of sources.