The AI was the easy part

Anyone who plays long strategy games knows the moment. You're on the couch, controller in hand, and you need one answer: which item a character needs, which choice keeps someone alive, how to unlock a class before the next chapter. So you pick up your phone, search, and land on a guide site that fights you. A video ad starts playing. A pop-up covers the text. The answer is somewhere in a guide split across 250 pages. By the time you find it, you've stopped playing.

I wanted to fix that for myself, so I built a small tool. It collects a full walkthrough from the web, strips out the clutter, and gets it into Google’s Gemini Notebook (formerly NotebookLM), where I can ask questions in plain English from my phone and get answers that cite the guide. It has nothing to do with my consulting work. It still taught me more about how organizations should approach AI knowledge tools than I expected a game project to.

I started by building the wrong thing

My first version was a complete custom app: its own database, its own search, its own AI layer, its own phone interface. With an AI coding assistant doing the typing, it worked.

Then I asked the obvious question: Why was I building and maintaining a search engine and a chat interface when Gemini Notebook already does both better than I ever will, and grounds every answer in the documents I give it? I set the custom app aside. The project became something much smaller: a way to get the right information into a tool that already exists.

That's the lesson I'd hand any organization looking at AI for its internal knowledge. The AI is the solved part. Plenty of products can read your documents and answer questions about them. None of them can decide which documents matter, get them out of the places they're stuck, and put them in shape to be useful. That work is yours, and it's most of the job.

The shape of the information decides the quality of the answers

Gemini Notebook limits how many sources one notebook can hold, and a single game guide ran past 250 pages. Loading them one at a time was never going to fit.

So the tool bundles them. Instead of 250 fragments, it produces seven organized books: one per major storyline, one for checklists and reference tables, one for combat strategy, each with its own table of contents. The answers got better, not just more convenient. When related chapters live in the same document, the AI can connect them. When they're scattered across hundreds of files, it answers from pieces.

Most organizations have the scattered version. The procedure lives in one file, the exception to it lives in an old email, and the current form lives in a shared drive nobody can navigate. Point an AI tool at that and you get confident answers built from fragments. Consolidating and organizing your source material isn't prep work you do before the AI project. It is the AI project.

Check what it grabbed before you trust it

Early on, the tool quietly stopped partway through a guide more than once. One run captured the first half of a game and missed the entire second half, and nothing about the output said so. So I added a report. Every run now lists how many pages it found, how many it captured, and the percentage that made it in. If the number isn't 100%, I know before I ask a single question.

Any organization feeding documents to an AI tool needs the same habit. The tool will answer from whatever you gave it, and it won't tell you what's missing. Someone has to know what should be in there and confirm that it is.

Decide where the automation stops

Everything up to the upload is automatic. The upload itself is a five-second drag and drop that I do by hand. That was deliberate. Automating it would have meant handing a script my Google login and inviting security lockouts on the account, all to save five seconds.

Full automation isn't the goal by default. The better question is where a person adds safety or judgment at a cost you barely notice. Keep the person there.

Swap the game for your organization

Replace the game guides with your employee handbook, your SOPs, your vendor documentation, your policy manual. The problem is the same: the information exists, it's scattered and hard to use, and the person who needs an answer needs it now, not after twenty minutes of searching.

The fix is the same too. Pick a tool that already handles the AI well. Spend your effort on what goes into it, how it's organized, how you check that it's complete, and where people stay involved. That part is less exciting than picking a tool. It's also where the results come from.