Free kit
Omid Sayfun
Senior software engineer in Amsterdam. 10+ years of keeping production software honest, most of them in fintech and payments. I've shipped many LLM systems to production, and I run several agent knowledge bases every day.
What you get
- 01
The audit prompt
Paste it into Claude Code, Codex, or Cursor opened in your knowledge base folder. Read-only. It returns a report card: pass, warn, or fail on each of the 14 checks, with file-level evidence and the smallest fix for everything it flags.
- 02
The 14 failure modes
Compounding wrong entries, duplicate concepts, silent supersession, notes that outlive the systems they describe, index drift, schema bloat, and the rest. Each with a symptom, a check, and a fix.
- 03
The prevention templates
A hardened schema file (CLAUDE.md) and a recurring lint checklist, so the audit findings stay fixed instead of coming back next month.
This is for anyone whose agent writes to a folder of markdown that later feeds its own answers: an Obsidian vault your agent maintains, a CLAUDE.md-driven project wiki, a team knowledge base fed by agents.
Every check comes from a failure practitioners actually reported, not from theory: the list is distilled from all 1,061 comments on Karpathy's llm-wiki gist. The full research is public: Knowledge bases for AI agents: 14 ways they rot.