There’s only one thing you really need to do to keep up with AI right now: create documentation.
Without documentation, AI can’t do meaningful work in your organization. With it, AI can already handle most of what gets done on a computer better than a human can. At TJ Digital, we’ve built our entire agency model around this premise, and we’ve watched dozens of clients go from treating AI as a novelty to running real revenue-driving work through it, all because they got their documentation right.
This is the abbreviated version of the process we run for clients. There are two types of documentation you need, and you can start building both today using Claude.
The two types of AI documentation every business needs
Most failed AI rollouts make the same mistake. They ask a model to know the company and know how to do the job, all in one chaotic prompt. That doesn’t work. The fix is to split the problem into two systems:
- A company knowledge base. The stable source material that tells the model who you are, what you sell, how you talk, what’s true, and what’s off limits.
- A library of skills. Reusable instructions for the recurring tasks the AI will actually perform: drafting proposals, writing emails, building briefs, onboarding clients, updating SOPs.
We call the knowledge base a brand ambassador when we build it for clients. The skills are exactly what they sound like: structured how-to documents the model loads when it needs to do that specific kind of work.
This split shows up across every major AI platform. Anthropic separates Projects (knowledge) from Agent Skills (procedures). OpenAI splits durable project guidance from skills that, in their words, “package repeatable processes.” Google Gemini does the same thing through Gems and Workspace context. The product names differ. The architecture is the same.
Why you should build your knowledge base in Claude
Use Claude for this part, even if your team is going to do its day-to-day work in ChatGPT or Gemini.
The reason is that documentation creation is a long-context, document-shaping task. Claude is currently the strongest tool for cleaning up messy material, holding multiple documents in working memory, and producing clean, structured output. Once the documents exist, you can connect them to any AI model. The knowledge base itself is portable. The act of building it isn’t.
I recommend Claude Cowork if you have access to it. Claude.ai works fine too.
How to build your company knowledge base
Start the conversation by telling Claude exactly what you’re doing. Use a prompt like this:
“I need your help putting together a company knowledge base. This will be a comprehensive set of documents covering information on my company and my industry. The goal is to create something I could give to AI models that would provide everything they could possibly need to work in and for my company. Your job is to collect all this information and structure it in a way where it’s easy and efficient to retrieve just the information you need for a given task.”
Then give it everything you have. Your website URL. Social media handles. Pitch decks. Brand guides. Pricing pages. Case studies. Founder bios. Customer personas. Approved claims. Disclaimers. Anything you’d hand to a new hire on day one.
Claude handles messy, unstructured input well, so don’t waste time cleaning it up first. End the prompt with: “Please ask me any questions that’ll help you do this better.” That gives the model room to surface gaps you didn’t know existed.
Separate public information from internal information
Here’s the pro tip that saves a lot of pain later: tell Claude to split the documents into two buckets. One for information that can appear in public-facing content. One for internal-only information.
If you skip this step, you’ll eventually catch your AI quoting internal pricing logic in a sales email or referencing an unreleased product in a blog post. Every major platform has access controls that respect this split, but only if your documents are structured to support it. The boundary has to live in the source material, not just in the prompt.
Add usage instructions
Once Claude has produced the first draft of the documents, ask it to write usage instructions. This should be a short description of:
- What each document contains
- When the model should use it
- Whether it’s internal-only or safe for public-facing content
This becomes the project instructions when you connect everything to an AI platform.
Where to host the documents
You have three main options. Each has a place.
| Hosting option | Best for | Trade-off |
| Local files | Quick prototyping, single-user setups | No team access, no version history |
| Google Drive | Teams already on Workspace; raw working files, decks, sheets | Less structured than a wiki |
| Notion | Curated wikis, page ownership, verified pages, governance | More upfront setup required |
For most companies, the right answer is Notion and Google Drive together, not one or the other. Put canonical, source-of-truth documents (process docs, SOPs, brand rules) in Notion, where ownership and freshness are visible. Keep raw files and collaborative working materials in Drive. Connect both to your AI platforms through projects, connectors, or MCP.
Connecting to multiple AI models
Once your documents are hosted, you can connect them to every major AI model. Create a project in Claude, ChatGPT, or Gemini. Drop the documents into the project’s knowledge base. Paste the usage instructions into the project instructions field.
You now have the same knowledge base running across every platform your team uses. That portability is the point.
How to build skills for every recurring task
The knowledge base is the first half. Skills are the second half.
A skill is a set of instructions any AI model can read. You’ll want one for every task the AI is going to handle for you. Drafting outreach emails. Building proposals. Writing blog posts. Updating SOPs. Anything you’d otherwise re-explain every time you opened a new chat.
The first skill to build: knowledge base maintenance
The first skill you build should be the one that keeps your knowledge base current.
In the same chat where you built your knowledge base, tell Claude:
“I now want you to make a skill for updating and maintaining this knowledge base. Whenever I give you new information about our company or industry, you should look through all of our current documents and determine if any of them need to be updated. If the information doesn’t fit in an existing document, you should create a new one.”
That’s it. Claude already has the context it needs from the conversation. Save the skill to your Claude account. Download a copy you can use on other AI models.
Now, any time new information surfaces (a pricing change, a new case study, a customer interview, a product update), you run it through this skill and your knowledge base updates itself.
Building skills for every other task
For every task the AI will handle, you want a skill. If you don’t have one yet, the move is simple:
- Start the task by telling Claude: “We’re going to do this task, and at the end, you’re going to create a skill for doing it in the future.”
- Walk through the task with the model, giving feedback as you go.
- When you’re done, tell Claude: “Please take all the feedback I’ve provided and everything you’ve learned, and update the skill to make it better for next time.”
The next time you need that task done, you tell Claude to use the skill. After the task, you ask it to update the skill again with anything new it learned.
Why this loop matters
This is the part most teams miss. The maintenance loop is the whole point.
Every time you use AI on a task, you’re either improving the knowledge base, improving a skill, or building a new one. OpenAI’s own Codex documentation puts it bluntly: when an agent makes repeated mistakes, you correct the guidance file and let the fix persist. Treat it as a feedback loop.
Companies that figure this out keep getting more out of AI as time goes on. Companies that don’t keep starting from zero in every chat.
The smarter the models get, the more important documentation becomes
There’s a common assumption that as AI models get more capable, the need for documentation will fade. The opposite is happening.
A more capable model can do more with what you give it. Give a weak model a strong knowledge base, and you’ll get decent work. Give a strong model the same knowledge base, and you’ll get work that outperforms most humans on the same task. Give a strong model nothing, and you’ll get confident-sounding output that’s generic, off-brand, and frequently wrong.
The model is replaceable. The documentation is the moat.
If you want AI to do meaningful work in your business, this is what you build first. If you’d like help building it for your company, reach out to TJ Digital and we’ll walk you through the same process we run for our clients.