How Do Product Managers Build an AI Research Knowledge Base?
For Product managers and UX researchers · Based on Karpathy Self-Improving AI Knowledge Base
// TL;DR
Product managers drown in research reports, user interview transcripts, competitive analyses, and meeting notes — most of which get read once and forgotten. The Karpathy Self-Improving AI Knowledge Base turns this scattered information into a structured, queryable Wiki maintained entirely by an AI librarian. Dump research artifacts into Raw, let the AI organize them by theme, and query the system before every product decision. The compounding loop means insights from Q1 research inform Q4 strategy automatically. Monthly health checks surface gaps in your research coverage before they become blind spots.
Why do product managers need a self-improving knowledge base?
Product managers generate and consume more research than almost any other role. User interviews, usability tests, competitive analyses, stakeholder meeting notes, analytics reports, industry articles — the volume is overwhelming. Most of this ends up in Google Drive folders or Confluence pages that nobody revisits.
The Karpathy AI Knowledge Base fixes this by creating a living research repository that the AI organizes, cross-links, and audits. You stop being the librarian and start being the decision-maker who queries a system that already knows everything your team has learned.
How do you structure a product knowledge base?
Create a domain folder like `product-research-kb` with Raw, Wiki, and Outputs subfolders. Write a Claude MD with focus themes tailored to your product — for example, "user onboarding friction," "competitive positioning," "feature prioritization frameworks," and "stakeholder communication."
Dump everything you have: user interview transcripts, research synthesis decks (converted to markdown), competitive analysis notes, product strategy docs, saved articles on PM methodology, and retrospective notes. Use the Obsidian web clipper to convert saved web articles to clean markdown. Don't organize anything — Raw is a junk drawer by design.
Run the Wiki build prompt. The AI will create topic pages for each major theme, cross-link related findings (e.g., connecting a user interview insight about onboarding friction to a competitive analysis noting a rival's superior onboarding flow), and build a searchable index.
How does querying the knowledge base improve product decisions?
Before your next sprint planning or roadmap review, query the knowledge base: "What does our research say about the biggest unmet needs in our onboarding flow?" The AI synthesizes findings across multiple research reports and user interviews, citing specific sources. The answer gets saved to Outputs, feeding the compounding loop.
Ask follow-up questions: "Based on everything in the Wiki, what are the three biggest gaps in our understanding of user retention?" The AI generates a gap report that doubles as a research planning tool — you now know exactly what to investigate next instead of guessing.
For team settings, update the Claude MD to attribute sources to team members. When a designer asks "What did we learn from the Q2 usability tests?", the AI pulls the relevant findings and credits the researcher who conducted them.
How do monthly health checks prevent research blind spots?
The seven-stage monthly audit is particularly powerful for product teams. The AI checks for contradictions between research findings (e.g., two studies with conflicting conclusions about a feature), identifies unsourced claims that might be assumptions rather than research-backed insights, and surfaces coverage gaps — areas where you have user data but no synthesis, or themes mentioned in strategy docs with no supporting research.
The health check report becomes an input for research planning. Instead of deciding what to research based on gut feel, you have a systematic audit showing exactly where your knowledge is thin.
Next step
Export your team's last three research reports as markdown files. Create your domain folder, write a Claude MD with your product's key themes, and dump the reports into Raw. Run the Wiki build and ask one question relevant to your current sprint. Even with just three reports, you'll see the AI surface connections you missed. Then commit to dumping every new piece of research into Raw going forward — the system compounds from there.
// FREQUENTLY ASKED QUESTIONS
Can multiple product team members contribute to the same knowledge base?
Yes. Each team member dumps material into the same Raw folder independently. Update the Claude MD to attribute sources to contributors so the AI tracks who provided each input. Designate one person as the system owner who runs ingestion passes and health checks to maintain integrity. Avoid concurrent ingestion runs — stagger contributions and process them in batches to prevent conflicts.
How do I handle user interview transcripts that are very long?
Drop the full transcript into Raw as a markdown file — the AI will extract key themes and findings during ingestion. For transcripts longer than 10,000 words, consider splitting them into separate files by interview or by section. The AI handles long documents but may lose fidelity on very long single files. Include participant identifiers (anonymized) so the AI can attribute findings to specific user segments in the Wiki.
Does this replace tools like Dovetail or EnjoyHQ for research management?
It serves a different purpose. Dedicated research tools offer tagging, highlight reels, and team collaboration features purpose-built for UX research workflows. The Karpathy knowledge base is a synthesis and query layer — it excels at connecting insights across all your sources and surfacing patterns you'd miss manually. Many product managers use both: a research tool for primary data management and the AI knowledge base for cross-cutting synthesis and strategic querying.