The hard part of a knowledge base is not putting files somewhere. It is keeping information findable, understandable, and reusable over time. I prefer to treat a knowledge base as a software system rather than an endlessly growing notes folder.
Move from folder thinking to a content model
Astro Content Collections work well as a static knowledge layer. Markdown remains the source format, but a schema requires every item to carry consistent metadata such as title, description, date, and tags. Navigation, search, RSS, archives, and related-content logic can then be derived from the same source of truth.
What Claude should and should not do
I use the model to help structure material, fill missing context, and identify duplication. I do not treat it as a source of facts. Primary material, citations, and important conclusions remain traceable to their original source of truth.
AI can reduce the cost of organizing information, but it cannot replace provenance.
A minimal workflow
- Put all source material in Git first so every change is reversible.
- Use schemas to constrain metadata instead of relying on naming habits.
- Let AI propose summaries and tags, then confirm them with a human or deterministic rules.
- Generate search indexes, RSS, sitemaps, and archives at build time.
- Add a database or server runtime only when a real interactive requirement appears.