Keeping help pages up to date is tedious work. New features, changed workflows, updated wording – until now, all of that had to be done by hand. We asked ourselves: What if AI didn't just answer questions, but maintained the knowledge itself? The result is our new help system – inspired by an idea that came out of Silicon Valley.
The Idea: Andrej Karpathy's "LLM Wiki"
Andrej Karpathy – co-founder of OpenAI and former head of AI at Tesla – published a concept in April 2026 that he called "LLM Wiki". The core idea: instead of having an AI search through all documents from scratch for every question (the classic RAG approach), you build a living knowledge system that maintains itself and improves over time.
"The tedious part of maintaining a knowledge base is not the reading or the thinking … it's the bookkeeping."
— Andrej Karpathy
The pattern consists of three layers:
The Three Layers of the LLM Wiki Pattern
Applied to Crazypatterns
- •Feature specs
- •Support exports
- •Screenshots
- •Legacy documentation
- •Knowledge pages
- •Help pages
- •Cross-references
- •Terminology / glossary
- •File format rules
- •Editorial standards
- •Lint rules & QA
- •Sync operations
What makes this special: the AI reads raw material, distills knowledge into interconnected pages, and keeps everything up to date – much like a librarian who doesn't just shelve books but also writes summaries and flags outdated entries.
What We Built From It
We found the idea immediately compelling – but our use case is different from a personal knowledge base. At Crazypatterns, it's about help pages that are derived directly from our code. When a feature changes, the documentation should automatically follow.
Inspired by this idea, we developed our own layer model for our code-to-help workflow:
From Code to Help Page – Our Layer Model
Unmodified source documents from the development process.
- Feature specs
- Support exports
- Screenshots
- Legacy documentation
AI-generated platform knowledge in a dedicated intermediate layer.
- Product knowledge
- Editorial rules
- Terminology
- Feature mapping
Finished help pages for the live website.
- Help pages
- JSON metadata
- Markdown body
Synced to the database, then live on the website.
Why Filesystem Over Database?
A deliberate design decision: the filesystem is our source of truth, not the database. Markdown files can be versioned (Git), reviewed (pull requests), and read and written directly by the AI. The database is just the runtime mirror – a fast cache for the live website.
How the AI Actually Works
Our system uses four operations that are loosely inspired by Karpathy's concept – but tailored to our workflow:
The 4 AI Operations in the Crazypatterns Workflow
raw/sources/* For readability, two steps (validate + audit) are combined into "Lint" here.
A concrete example: when we change the download process, a developer drops the new spec into raw/sources/.
The AI reads the spec, updates the knowledge layer, and then adjusts the relevant help page for downloading patterns – in all supported languages.
Ensuring Quality: From Editorial Standards to Automated QA
AI-generated content needs clear guardrails. That's why we defined a clear editorial standard that applies to every help page:
- One page = one user goal. No catch-all pages, no mixed topics.
- The answer starts at the top. What is this about? Who is it for? What's the outcome?
- Real step-by-step instructions. Numbered steps, clear actions, expected results.
- No marketing fluff. Help is help, not advertising.
- Links must be self-explanatory. No "click here" – instead, "Reset your password."
QA Pipeline
From Markdown file to verified live page
- AI proposes, a human decides
- For now, no automatic publishing without manual approval
Our QA suite automatically checks four areas:
| Check | What's verified? | Method |
|---|---|---|
| Links & Anchors | No .md links in the frontend, no wrong locales, all anchors present | Link resolution |
| Content Structure | No swallowed paragraphs, clean heading hierarchy | Structure validation |
| Special Elements | Info boxes not empty, formulas not truncated, tables properly styled | CSS/JS inspection |
| Image Sequences | No bullets next to images, no side-by-side errors, no broken images | Asset mirroring |
The Fact Checker
A key component of our pipeline is the AI-powered fact checker. This agent regularly inspects our code and compares the help pages against the actual state of the platform.
Here's how it works:
- Code inspection: The agent reads relevant code sections and compares them against the descriptions in the help pages.
- Discrepancy detection: If a help page says "Click the blue button" but the button is now green, that gets flagged.
- Language parity: Automated checks to ensure DE and EN pages are in sync.
- Continuous monitoring: The check runs automatically – not just when someone remembers to do it.
Fact Checker – Code ↔ Help Page
The AI fact checker automatically compares the platform's actual implementation against the official documentation.
form has 3 fields.
Reads relevant code sections
Flags mismatches (e.g., color, field count)
Checks if DE & EN are in sync
"Fill in 4 fields"
- Concrete change suggestions for help page text.
- Human review: The AI proposes, a human decides.
What This Means for You
For you as a Crazypatterns user, one thing changes above all: the help becomes more reliable.
- Help pages are updated faster when something changes on the platform.
- Fewer outdated screenshots and incorrect descriptions.
- Better consistency across languages.
- Clear, task-based structure instead of long walls of text.
We believe that good help isn't an afterthought. It's part of the product. And with the right tooling, we can maintain it the way we maintain our code: carefully, transparently, and continuously.
