Build a Team UX Research Knowledge Base With AI
For product managers · Based on Karpathy Self-Improving AI Knowledge Base
// TL;DR
Product managers can build a self-improving AI knowledge base to centralise UX research methods, past research reports, and conference notes that would otherwise be lost across drives and Slack. An AI librarian compiles a cross-linked Wiki on qualitative methods, synthesis techniques, and stakeholder communication. You query it to check what your team recommends for rapid generative research, save the answers back, and let monthly health checks surface gaps — like missing coverage on remote or async research — so you commission new content deliberately instead of by accident.
Why do product teams lose their research knowledge?
Product teams generate enormous research value — usability studies, generative interviews, synthesis workshops, conference talk notes — then bury it in shared drives nobody revisits. New PMs re-run studies that already exist, and hard-won methodological lessons evaporate when someone leaves. The Karpathy self-improving AI knowledge base solves this by appointing an AI as your team's research librarian. Everyone dumps material into one Raw folder, and the AI builds a queryable, cross-linked Wiki of your team's collective research knowledge.
How do I structure it for UX research?
Create a second brain container and a domain folder called `ux-research-kb` with Raw, Wiki, and Outputs subfolders plus a Claude MD. Set three themed focus areas in the Claude MD: qualitative research methods, synthesis techniques, and communicating findings to stakeholders. These tune the AI to deepen exactly the areas your team relies on.
One important adjustment: the default Karpathy architecture assumes solo use. For a team, update the Claude MD to acknowledge collaborative inputs and attribute sources to specific team members. That way, when the Wiki cites a synthesis technique, it credits the researcher who documented it — preserving provenance across the team.
How do I get the research into the system?
Dump all your team's past research reports into Raw as markdown, along with saved articles and conference talk notes. Use the Obsidian web clipper for web articles and paste transcripts directly into the AI chat to be saved as MD files. Don't reorganise anything — Raw stays a junk drawer.
Then run the Wiki build: point the AI at the folder and prompt it to read Raw and compile the Wiki per the Claude MD, index first, one file per major topic, cross-linked. In about 30 minutes you get topic pages summarising your team's methods, with discovered connections between studies and a searchable index.
How does querying surface research gaps?
Ask the AI: 'What methods does our knowledge base currently recommend for rapid generative research?' It answers from your team's own reports, citing sources. Save the report to Outputs. Then ask for the biggest gaps — the report will likely surface that your base contains no content on remote or async research methods, or that stakeholder communication is thinly covered. Use this to commission and ingest new content deliberately, instead of discovering the gap mid-project.
Run the monthly seven-stage health check to catch contradictions between studies, broken references, stale reports over 90 days, and coverage gaps. The AI drafts new article candidates and, in interactive mode, you choose which to action. Over time, this becomes onboarding gold — a new researcher can query the base instead of interrupting the team.
Keep the discipline: never edit the Wiki by hand, always save query outputs back, and maintain the Change Log so ingestion stays accurate as multiple people add material.
Next step: Set up your `ux-research-kb` folder, update the Claude MD for team attribution, and have each researcher drop their most recent report into Raw before your first Wiki build this week.
// FREQUENTLY ASKED QUESTIONS
How do I adapt the system for a whole team rather than one person?
Update the Claude MD to acknowledge collaborative inputs and attribute sources to specific team members, since the default architecture assumes solo use. Add attribution rules so the AI credits who contributed each source. Maintain a shared Change Log so the AI tracks what's new as multiple people add material, and stagger health checks to manage credits.
Can new team members use this for onboarding?
Yes. A new researcher can query the knowledge base to learn your team's recommended methods and synthesis techniques instead of interrupting colleagues. Because the AI cites sources from your actual past reports, they get context-rich, team-specific answers. Point them at the folder and let them ask questions, then save useful outputs back for the next person.
How does this stop us re-running research that already exists?
Before commissioning new research, query the knowledge base for what your team already recommends on that topic. The AI surfaces existing methods and reports with sources. If it finds nothing, the gap report tells you deliberately — so you commission genuinely new work rather than accidentally duplicating a study buried in a shared drive.