The site is the system — how a living knowledge base is built
A SharePoint knowledge base is no longer a folder of documents. It is a site whose document library, submissions folder, and context file together form the substrate that AI Skills act on.
3 things to walk away with
Substrate
A KB is a SharePoint site with libraries, lists, and a context file working as one organism.
Intake
Submissions are raw drop-offs; AI skills decide how and when to promote them into the KB.
Governance
The context file is the rulebook — it tells every agent session how to behave.
The Knowledge Base Site
The demo opens with a SharePoint site that looks ordinary — but every piece of it has a purpose. It hosts a document library of markdown files (the canonical knowledge), a submissions folder for raw intake, a gaps list, an expert directory, and a context file that defines the rules. Together these form a living knowledge base, not a static document dump.
A living KB is a site shape, not a single document store. The library, lists, and context file work together so AI can read, write, and reason about the whole.
Treat the site as the unit of design. When you scaffold a new KB, you are not creating files — you are wiring up the surfaces an agent will read from and write back to. Every workflow in this video, from question-answering to ingestion to admin reports, depends on this same shape.
A SharePoint knowledge base is just a document library. In a living KB, the document library is one of several surfaces — gaps lists, expert directories, and the context file are first-class parts of the system.
The Document Library — markdown as canonical knowledge
The library holds markdown files like customer-onboarding.md and vendor-contracts.md. Crucially, none were written by humans. AI skills generated them by reading raw inputs — meeting transcripts, emails, decisions — and writing back structured markdown that captures not just the facts, but the rationale and nuance behind them.
Markdown is the agent-friendly format of choice. It is plain text the AI can read and rewrite, while staying human-reviewable in the browser.
Each file carries a last-verified date in its metadata. The agent reads this on every query and uses it to decide whether the answer is still fresh, stale, or worth flagging. That single piece of metadata is what makes the KB feel alive.
You should paste your meeting notes directly into the library. The system works because raw inputs are processed into curated markdown — the agent decides what belongs in the canonical entry, not the user.
The Submissions Folder — raw intake without commitment
The submissions folder is a separate location where anyone can drop content — meeting recordings, email threads, transcripts, decks, even random PDFs. Nothing is added to the KB yet. Submissions are just queued raw material waiting to be processed.
An approver (or an automated approval flow) then triggers the ingest skill. The skill reads the dropped files, extracts what belongs in the KB, decides which existing knowledge files to update, and even resolves outstanding knowledge gaps that the new content happens to answer.
The submission folder decouples capture from curation. People can drop anything; the AI decides what — if anything — graduates into canonical knowledge.
Submissions are auto-ingested the moment they land. They sit and wait — ingestion runs only when an approver triggers it, so you stay in control of what enters the KB.
The Context File — the rulebook every agent session reads
The context file is a markdown document the agent loads at the start of every session. It defines what the KB is, how knowledge should be stored, where reports go, how to name files, how to handle decay, and the rules for answering questions (e.g. always cite the source, always check the expert directory before assigning a gap).
Behaviors you see in the demo — flagging stale entries, logging verification items automatically, refusing to create duplicate gaps — all come from this single file. Edit the context file and the whole site changes behavior.
The context file is the single highest-leverage surface in the system. Most of what looks like model intelligence is actually rules the agent reads on every run.
Each AI Skill carries its own behavior policy. Skills are the verbs (ingest, log, route, report); the context file is the grammar that ties them together for a given site.
Flashcards — Foundations
01 · Foundations
Knowledge Base Site
tap to reveal →
A SharePoint site assembled from a document library, submissions folder, gaps list, expert directory, and context file — the substrate that AI Skills operate on as one organism.
← tap to flip back
01 · Foundations
Document Library
tap to reveal →
The library of canonical markdown knowledge files. Files are produced by AI skills (not pasted by humans) and carry a last-verified date the agent reads on every query.
← tap to flip back
01 · Foundations
Submissions Folder
tap to reveal →
A staging location where users drop raw inputs (transcripts, emails, decks). Nothing enters the KB until an approver triggers the ingest skill — capture is decoupled from curation.
← tap to flip back
01 · Foundations
Context File
tap to reveal →
A markdown rulebook the agent loads at the start of every session. Defines storage conventions, decay handling, source-citation rules, and gap-routing — most "smart" behavior lives here.