How to Use Gemini: The Prompts That Fail in Google Docs and Sheets, and the Fixed Versions

When you attach AI to your work, the time that leaks most is not the model’s performance — it is the round trip. Open the document, copy the part you need, paste it into the chatbot window, move the result back, fix the broken formatting. Two or three minutes each time, but twenty times a day is an hour.

What Gemini inside Google Workspace solves is exactly that round trip. So this article organises five scenarios not by “how smart is it” but by “which round trip disappears”.

One thing to note first. In a shared document, Gemini reads only what you already have access to. Anything you write into a team document is immediately visible to everyone else, so the safe habit is to draft in a personal document and move only the final version.

Scenario 1: Cleaning up meeting notes

The source: notes typed during the meeting, or an automatic transcript. Sentences trail off and speakers are mixed together.

• The prompt that fails: Clean up these meeting notes
• The fixed prompt: Turn these notes into meeting minutes. (1) a three-line summary (2) decisions (only what was actually agreed) (3) open items (4) action items as a table (task / owner / due date). Do not invent owners or due dates that are not in the source — leave them as "TBD".

Always include the sentence telling it not to invent things. The dangerous error in meeting minutes is not an awkward sentence; it is something that was merely discussed being promoted to a decision, or a deadline nobody mentioned being filled in plausibly.

Scenario 2: Spreadsheet formulas and cleanup

Building formulas from plain language is the biggest payoff on the spreadsheet side.

• Write a formula that sums column D for rows where column B is "Done"
• Show me how to find and flag duplicate values in column A
• This date column came in as text. Give me a formula that converts it to a date format

The time spent looking up the argument order of SUMIF disappears. In exchange, you always verify. The method is simple: compute a few rows by hand and compare, or change one condition and see whether the number moves the way you expected. Formulas are unlike prose — when they are wrong they still emit plausible numbers.

Scenario 3: Summaries and drafts from long documents

A summary prompt only becomes usable once you fix three things: the audience, the length and the format.

• The prompt that fails: Summarise this
• The fixed prompt: Summarise this document as if reporting to a manager who is seeing this project for the first time. Five bullets maximum, roughly one line each. If a number appears, it must be included.

Drop any one of the three and the result wobbles. “Numbers must be included” in particular prevents the common problem of figures evaporating wholesale from a summary.

Scenario 4: Email drafts and adjusting the tone of replies

In mail it is specifying tone, rather than generating sentences, that earns its keep. The advantage is that it can read the received thread as it stands.

• External: Draft a reply to this thread. Polite, but commit firmly to the schedule. Five sentences maximum.
• Internal: Same content, in the tone I'd post to a team channel. Skip the greeting, conclusion first.

Keeping the same content in two tones lets you pick whichever fits the situation. Just do not send the draft as it is. AI mail tends to run a sentence or two long and to be excessively polite. Cutting it in half is part of the same set.

Scenario 5: Building a research table

A prompt like Put the pricing plans of five competing services into a table produces the fastest results and is the most dangerous task.

Tables hate empty cells. If a value is unknown, a plausible number may be filled in, and the credibility a table format lends means people question it less. Figures that would hurt if wrong — prices, launch dates, market share — must be re-checked against the original source.

The safe way to use it is to invert the order. First build the skeleton with Just list the items (columns) I should compare, then fill the values in yourself from the official pages. The AI supplies the structure of the research and a person is accountable for the facts.

⚠️ Workspace AI features differ in scope by account type (personal / business / education) and by plan, and the policies change often. Check the current terms in the admin console and the official documentation before rolling anything out.

Frequently asked questions

Does it work on a free Google account?

Features and limits differ by account type. What is available on a free personal account and what unlocks on a paid Workspace plan are separate, and that boundary moves with policy changes fairly often. The fastest check is to open a document with the account you will actually use and see whether the feature appears.

Is it safe to put company documents in?

Business Workspace plans generally guarantee contractually that organisational data is not used to train models, but whether that applies to you depends on your organisation’s plan and your administrator’s settings. Before handling confidential documents, check the data policy in the admin console and your company’s own security rules first.

Does it handle Korean documents well?

Summaries and drafting are good enough for real work. The level of formality is not uniform, though, so internal and external tones can get mixed, and long Korean text inside a table wraps awkwardly. Specifying the tone in the prompt resolves most of it.

How is this different from pasting into ChatGPT?

The difference is less in output quality than in how much handling it takes. AI inside the document reads the file and thread that are already open and writes the result in place, so copying, pasting and formatting repair disappear. Conversely, gathering material from several services in one place and refining it over a long conversation is easier in a separate chatbot.