Turn Google Docs Into a Gemini Knowledge Base

A question I get a fair bit at work has nothing to do with Salesforce: how do I stop losing my best AI prompts and chats? People build up a handful of prompts that really work, or they have one great Gemini conversation, and a week later they're scrolling through a wall of chat history trying to find it, or retyping the whole thing from memory.

I took a run at this a while ago. I built a small Chrome extension for my team, the Ateko Gemini Sidekick. The first version sorted your Gemini chats into folders and kept a library of reusable prompts. Then a Gemini UI change wiped out a year of my neatly organized chats overnight, so this year I rebuilt it to focus on the prompt library and let the folders go. It's handy. You can store a good prompt and drop it back in with a click.

But it clearly didn't fix the real problem, because the question kept coming up long after everyone had the tool. Storing a prompt and pasting it back in is not the same as giving Gemini the knowledge behind that prompt. One hands over a paragraph of instructions. The other teaches the model what you know.

Over in Cursor, where I do most of my real work (mostly Claude Opus and Sonnet, with a few other models mixed in), that second part is already handled. Cursor runs on a healthy stable of skills, rules, and personas: markdown files I share with the team that Cursor feeds to the model on every request. One is a beefy document of AI writing tells. I never re-explain any of it. The file is there, and the model follows it.

I wanted the same thing in Gemini. It turns out you can get most of the way there with nothing but a Google Doc.

Why a Gem isn't the answer

A Gem is Gemini's version of a saved custom assistant. You can attach knowledge files to it, and that works fine, up to a point. The point is ten files. That's the cap. You also get one instructions box for the whole Gem, so there's no clean way to tell it what each file is or when to use which one. For a small, fixed set of reference docs, a Gem does the job. For a knowledge base you want to keep growing, it runs out of room fast.

Point Gemini at a Google Doc

The whole setup takes about five minutes.

1. Go to Google Docs and start a new doc. Name it "My Writing Rules."
2. Paste in a few rules to start with. Steal these five, then add your own as you go:
   • No em dashes.
   • Never use these words: delve, tapestry, leverage, robust, seamless, pivotal.
   • No "it's not X, it's Y" reframes (not "it's not a tool, it's a teammate," just say what it is).
   • Don't stack lists of three for rhythm (one is fine, three in a row is a tell).
   • Cut the hedging: "it's worth noting," "it's important to note," "arguably." Say it straight.
3. Open a new Gemini chat.
4. Click the plus button next to the prompt box, choose Add from Drive, and pick your doc.
5. Write your prompt like normal, then add: "check it against my writing rules doc."
6. Want to change the rules later? Edit the doc. The next answer uses the new version, and you never re-upload a thing.

One note on step 4: you can also type @ to pull in a Drive file by name, but in my testing that was hit and miss (I asked for two docs by name and it grabbed one). The plus button and Add from Drive is the path I trust, because you can see exactly what's attached.

A word on how deep those rules should go. The flat bans stay one-liners forever: you just keep adding words. The judgement calls are the ones that get sharper when you hand the model an example, like I did with the reframe and the rule of three above. That's the whole reason my own rules doc is thousands of words and not five bullets. It isn't more rules, it's more examples of what I mean.

Before I trusted this, I ran a simple test to prove it holds up. This next part is me kicking the tires, not steps you need to repeat. I made a Doc with two rules ("never use the word delve", "flag every em dash") and a made-up fact ("the project codeword is BLUEHERON"). Then I attached it and pushed on it:

1. I pasted a sentence that used the word "delve" and had an em dash in it, and asked Gemini to check it against the doc. It caught both, cited the doc, and offered a rewrite.
2. I asked what the codeword was. BLUEHERON, straight from the doc.
3. Then I went off-topic for fifteen messages. When I came back and asked again, it still had both rules and the codeword. It had even been tagging its own em dashes with "[FLAG: em dash]" as it went, which was both useful and a little ridiculous.
4. I edited the Doc to change the codeword to REDFOX. When I asked again, it said REDFOX. No re-uploading.

That last step is what makes this worth doing. The Doc is live. Edit it and Gemini sees the change, and not just in your next chat. A conversation that already has the Doc attached picks up the edit too. So you can build the knowledge base on the fly, mid-conversation: notice a rule you want, add it to the Doc, and keep going. Same loop Cursor gives me with its rules files, without the repo.

Three docs worth building

Same approach for each one: make the Doc, keep feeding it, attach it with Add from Drive, and let it steer the chat.

1. Your writing rules

This is the one that pays off the most, because it lets you do the work once. You can spend real time researching what good looks like, get it right, drop it in a Doc, and then haul that effort into every chat from then on.

Mine is exactly that. It's a roughly 4,300-word reference on AI writing tells: the words, phrases, and patterns that make text read like a machine wrote it. I pulled it together from Reddit threads, journalism, academic papers, and editors complaining about the AI submissions landing in their inboxes. It covers the obvious stuff (the em dash, "delve," the rule of three) and the subtle stuff (hedging, fake balance, the influencer voice). Attach it to a chat and Gemini checks a draft against all of it. I don't re-explain my taste every time, and I didn't have to remember 4,300 words of it either. The Doc remembers for me.

You don't need mine. Start your own: words to ban, the em dash ban, notes on how you want things to sound. Every time a draft slips something past you, add the rule. It compounds.

2. Your best prompts

Make a Doc called My Best Prompts. Paste in the prompts that work, each with a line on when to use it. This is where a paste-library and a Doc part ways: pasting hands over one prompt, while attaching the Doc gives Gemini the whole set to pick from. Feed it when a prompt lands, fix it when one flops, cut the dead ones. Then attach it and say "use my editing prompt on this," or "pick the best prompt from that doc for what I'm about to do." You stop hunting, and the model does the matching.

3. Things you've learned from your chats

Make a Doc called Things I've Learned. It's the running log of the good stuff from chats worth keeping, and it's the one people skip. Here's the prompt I use, borrowed from my project work: before you leave a chat that mattered, ask "Review this chat and write up what's worth keeping: the decision, the reasons behind it, and the details I gave you that shaped it." Paste the result into the Doc with a date on it.

Save the context, not just the answer, because in six months the reasons are what you'll need. Then attach it when you come back to the topic, and the new chat starts where the old one left off instead of from zero.

Where it breaks

A few things to watch for, because this isn't magic.

Use the plus button, not @. The @ keyword search didn't reliably find my files. Add from Drive lets you see exactly what's attached.

Long chats lose the thread. Google advertises a context window of up to a million tokens, but the chat app works with much less than that at any one moment. On the free tier it's about 32,000 tokens, and even paid users have documented Gemini forgetting instructions after roughly 25 to 30 messages in a single chat. Your attached Doc can drop out of that active window without warning. My fifteen-message test held, but on a longer session, when the answers start ignoring your rules, re-attach the Doc or ask for a recap to pull it back into focus. That is also why this setup is so hard to break. Because the knowledge lives in the Doc and not the chat, a forgetful or cluttered chat is never a dead end. Start a fresh conversation, attach the Doc, and you are back to the same baseline in seconds.

It finds, it doesn't memorize. Gemini pulls the parts of the Doc that match your question. If nothing matches, it can fall back on its training and answer confidently anyway. Keep the Doc clear and well labelled so the right parts get found.

Keep it short and dated. A bloated, stale Doc is worse than a short one, because you'll trust it. Date your entries and cut what's wrong. If you're wrangling dozens of documents, that's a NotebookLM job, not a single Doc.

Don't put client or private data in a personal Gemini or Drive. This is the one that matters. A personal AI account is not the place for client information, customer records, or anything under an NDA. Keep your knowledge base to your own prompts, your own style, your own non-sensitive notes. Client work belongs in the tools and accounts your company has vetted for it.

The Doc outlasts the chat

The chat is where the work happens. The Doc is what's left when the chat is gone, and it's what makes the next one smarter. Cursor taught me that with its rules files. Gemini does the same thing, as long as you hand it a Google Doc instead of making it guess. Build one, keep feeding it, and stop starting from scratch.