How Product Managers Build a Self-Improving Research Base
For Product managers and UX researchers · Based on Karpathy Self-Improving AI Knowledge Base
// TL;DR
Product managers drown in user research reports, competitive analyses, strategy docs, and stakeholder feedback — most of which gets read once and forgotten. The Karpathy Self-Improving AI Knowledge Base turns this scattered knowledge into a living, queryable system. Dump research reports, interview transcripts, feature specs, and saved articles into Raw. Claude builds a cross-linked Wiki covering your product domain. Query it before roadmap reviews, stakeholder meetings, or strategy sessions. Monthly health checks surface research gaps and stale insights, keeping your product knowledge current and compounding.
Why Do Product Managers Need a Self-Improving Knowledge Base?
Product management generates enormous volumes of information: user interviews, usability test reports, analytics summaries, competitive analyses, feature specs, strategy memos, and stakeholder feedback. Most of this gets filed in Confluence or Google Docs, read once, and never synthesised.
The result: product decisions get made on whatever the PM remembers or can find in the moment, not on the full body of accumulated evidence.
The Karpathy Self-Improving AI Knowledge Base fixes this. The AI reads everything, organises it into a cross-linked Wiki, and gives you query access to your entire knowledge corpus. The compounding loop means every research report you add and every question you ask makes the next answer more comprehensive.
How Do You Structure the Knowledge Base for Product Work?
Create a domain folder like `product-kb` with Raw, Wiki, and Outputs subfolders and a Claude MD. Define themed focus areas that match your product domain. For a B2B SaaS PM, these might be: user onboarding patterns, enterprise buyer psychology, and competitive positioning. For a consumer product PM: engagement loops, retention mechanics, and growth channels.
Dump everything into Raw: past user research reports (exported as markdown), interview transcripts, competitive analysis docs, feature specs, strategy presentations (as text), analytics summaries, and relevant articles. The Obsidian web clipper handles articles; for Google Docs, export as markdown or plain text.
For team use, update the Claude MD to attribute sources to team members. Anyone can dump into Raw. One person triggers health checks monthly.
How Does the Knowledge Base Help Before Big Product Decisions?
Before a roadmap review, query: Based on all user research in the Wiki, what are the three strongest signals about what users need next? The AI reads the index, pulls relevant Wiki entries, and generates a sourced briefing saved to Outputs.
Before a stakeholder meeting, query: What does our knowledge base say about the competitive landscape for [feature area]? You get a synthesised answer pulling from multiple competitive analyses you've collected over months.
Before a strategy session, run: What are the biggest gaps in our current product knowledge? The gap report tells you exactly what research to commission next. This transforms your research practice from reactive to systematic.
How Do Monthly Health Checks Prevent Research Debt?
Product teams accumulate research debt — outdated findings, contradictory data from different studies, and topics that were never properly covered. The seven-stage health check catches all of this.
The AI flags contradictions (e.g., two studies with conflicting findings on user preferences), identifies unsourced claims, finds coverage gaps (e.g., extensive onboarding research but nothing on churn reasons), and proposes new articles. The PM reviews findings in interactive mode and decides which to action.
This monthly ritual takes 20 to 30 minutes and prevents the knowledge base from slowly degrading. Over six months, it builds a product knowledge asset that new team members can query from day one, dramatically reducing onboarding time.
Start today: export your last ten user research reports as markdown, dump them into Raw, and run your first Wiki build.
// FREQUENTLY ASKED QUESTIONS
Can my whole product team contribute to the knowledge base?
Yes. Update the Claude MD to acknowledge multiple contributors and attribute sources to team members. Anyone on the team can dump content into Raw — research reports, interview notes, competitive analyses. One person should own the monthly health check to ensure consistency. The Wiki remains AI-only territory regardless of team size. Use a shared cloud folder (Dropbox, Google Drive) for access.
How do I handle user research with sensitive participant data?
Anonymise research data before adding it to Raw. Replace participant names with identifiers, remove identifying details, and strip PII. The knowledge base runs on local files, so it's as secure as your file system, but anonymisation is good practice regardless. The AI doesn't need participant identities to extract useful patterns and insights from research findings.
Will the knowledge base help me find contradictions in user research?
Yes — this is stage one of the monthly health check. The AI audits all Wiki articles for contradictions and inconsistent data. If two research reports reach different conclusions about user behavior, the health check flags it with citations to both sources. You then decide whether to commission new research, weight one source higher, or note the contradiction explicitly in the Wiki.