AI Knowledge Base for Product Managers

If your product work is AI-centric, you know the pattern. You start a task by crafting a prompt, then compose the context to get from generic results to something useful. Company brief, target audience, goals for this quarter. All relevant, all necessary.

The problem is that composing this context gets repetitive fast. You're providing the same background every session. The same strategic priorities. The same constraints. The actual thinking hasn't changed — but you're rewriting it every time.

The model isn't the bottleneck. The missing context is. In an AI-based workflow, your local file system with markdown documents becomes the natural place to store this context and build your knowledge base.

Why not just use Confluence, Notion, or Google Docs?

Fair question. Your knowledge already lives somewhere — probably across several tools that offer collaboration, commenting, and rich UIs. That's genuinely useful for working with other people.

But it has a key drawback. Those documents are designed to be read by humans. We can easily focus on what matters and ignore the rest. An AI can't do that as well. The more bloated your context gets, the lower the quality of the output. What matters to your colleague right now might not matter to you for this task.

What I like about a local markdown knowledge base is that you can craft context in a way that works for both humans and AI. You're building a curated briefing, not mirroring your company wiki. Every file earns its place by helping the AI understand your product right now. Nothing is there by accident.

There's a second benefit: when curated well, the knowledge base is tool-independent. If you switch from one LLM provider to another, your core asset stays in a portable, readable format. No exports, no migration headaches.

How do you keep it from getting messy?

A knowledge base can get chaotic quickly. Where do you put the company strategy? Where do your personas go? What about release notes? Just piling everything into an endless stream of documents isn't the answer.

The context you compose is dynamic. The same task can draw from completely different inputs. Creating a B2B persona might use job descriptions one time and interview transcripts the next. Same result, different source material. So the structure needs to be flexible enough to handle that.

How we structured it

We took a few design principles and put them at the center. We wanted a structure that:

• Doesn't make strong assumptions about which frameworks you use in your product work

• Still gives PMs clear orientation on where to find and put things

• Stays clean over time — no slowly decaying board where nobody archives anything

• Has a starting point for messy, unprocessed work so you can get going without thinking about structure

• Uses AI to help keep the knowledge base organized

This became PO-1 — an open-source file structure that ships with a cleanup skill and a human/AI-readable guide that explains where specific artifacts belong as you generate them over time.

The folders map to how product people actually think. You don't need to decide whether a competitive analysis is a "project" or a "resource." It goes in research. A retro summary goes in team. A half-baked feature idea goes in workbench until you know what it is.

Inside each folder there's a temporal structure — current, planning, and archive:

When a quarter ends or a project ships, the documents move to archive. Your active folders stay focused on what's current.

The workbench: start here

If you try PO-1, start with the workbench. Not the structure.

Every organizational system I've abandoned had the same failure mode: I had an idea, opened the system, and spent more time figuring out where to put the thing than capturing it. The workbench removes that friction. Drop anything in there. The question changes from "where does this go?" to "is this worth writing down?" The sorting comes later — or the AI handles it for you.

The cleanup skill

PO-1 ships with a cleanup skill at .claude/skills/cleanup/. When you tell the AI to organize your workbench, it looks at each file's content and age and presents a recommendation table: which files to move, where they should go, and why.

It flags items untouched for 30+ days. It catches strategy docs referencing a quarter that already ended. It spots drafts that have matured enough to move into their proper folder. Nothing moves until you confirm.

The system doesn't just tolerate mess — it has a built-in way to turn scattered information into something organized when you're ready.

Get it and try it

PO-1 is free and open source: github.com/productbench/po-1

If you know git:

Open the folder in Claude Code, Cursor, or any editor. Drop in your product overview. Done.

Never used git? Go to the repository page, click the green "Code" button, select "Download ZIP." Unzip it and open the folder. All the structure and templates are there.

A good first step: think about the one piece of context you've typed into an AI prompt three times this week. Write it down in a markdown file. Drop it in the workbench. That's your knowledge base started.

Ready to go deeper with AI for product management? Join our upcoming "Master AI for Product Management" course.