Schema-Driven Maintenance
An LLM wiki becomes more reliable when the assistant follows an explicit schema instead of behaving like a generic chatbot.
The model still matters. But the operating rules matter just as much.
What the schema does
A good schema defines:
- directory structure
- page types
- naming rules
- frontmatter conventions
- ingest behavior
- query behavior
- lint behavior
- logging expectations
Without that, the assistant tends to produce one-off output. With it, the assistant can act more like a maintainer.
The important shift
Schema turns the job from:
- answering loosely
into:
- maintaining knowledge deliberately
That is a big difference.
Three operational modes
1. Ingest
When a new source appears, the assistant should:
- read it
- summarize it
- update relevant concept or topic pages
- update the index
- append to the log
2. Query
When asked a question, the assistant should:
- navigate the compiled layer first
- read the most relevant maintained pages
- synthesize an answer
- optionally file durable answers back into the wiki
3. Lint
Periodically, the assistant should check for:
- contradictions
- stale claims
- weak cross-links
- orphan pages
- missing concepts that deserve their own page
Why this matters
A smart model without structure often feels impressive but unstable. A structured maintenance workflow is less flashy, but it compounds better.
It creates a system that can be inspected, corrected, and improved over time.