Watch the demo and you see the agent log gaps, ingest knowledge, find experts, and reason about decay — these are AI Skills, the verbs the site exposes to every user session.
3 things to walk away with
Skills as verbs
The KB exposes four skills users can call by name; the agent picks which to invoke.
Decay is automatic
Every answer ships with a freshness check — stale entries spawn verification tasks on their own.
Why over what
Ingestion captures rationale, not just numbers — so questions like "why 99.9%" can be answered later.
The Four Skills — verbs the KB exposes
The demo KB ships with four AI Skills: log gaps, ingest knowledge, find experts, and generate admin reports. Each is a discrete capability the agent can invoke during a session. Users do not call them by name — they ask natural questions, and the agent decides which skills to chain together to satisfy the request.
Skills are the verbs of the site. The context file (the grammar) tells the agent when to use which verb on which surface.
Because skills are first-class building blocks, you can add or remove them per site. A legal KB might add a "redaction" skill; a sales KB might add "score-lead". The four shown here are the minimum set for a self-maintaining KB.
Skills are SharePoint plugins or apps the user manually activates. They are agent-callable capabilities — the agent invokes them based on the user's question and the context file's rules.
Asking Questions — the agent grounds answers in the library
The first demo question — "what is our policy on vendor contract SLA requirements?" — triggers the agent to scan the document library, find the relevant entry, and synthesize an answer. It does not invent. It cites the markdown file the answer came from.
What looks like simple Q&A is actually the entry point that unlocks the rest of the system. The agent also checks the file's last-verified date, notices it is over 90 days old, and silently invokes the log gaps skill to spawn a verification task — without the user asking.
A user question can fire multiple skills in sequence. The answer is just one output; the side effects (gap logging, task routing) are equally important.
The agent answers from a generic LLM model. It is grounded specifically on the document library, and the answer always traces back to a markdown file in the site.
Knowledge Decay — every answer has an expiration check
Each markdown file in the library records a last-verified date. The context file defines decay rules: anything over 90 days is "potentially stale" and should be flagged. When the agent answers from a stale file, it spawns a verification gap and assigns it to the expert who owns that domain.
In the demo, asking the SLA question triggers a verification item routed to Jordan Lee. Asking a second question about the same domain does not create a duplicate gap — the context file rule "do not log duplicate gaps" prevents it. The system is opinionated about not wasting attention.
Decay handling is the difference between a static wiki and a living KB. Every answer carries a freshness signal, and stale ones generate work.
Decay is a passive flag the user must notice. The agent acts on decay automatically — it creates tasks, routes them, and tracks resolution without prompting.
The Ingest Knowledge Skill — promote raw input into curated entries
When a user pastes "Hey, I can answer that — standard SLA is tier one…" into the chat, the agent detects this is new knowledge, not just conversation. The ingest knowledge skill fires: it parses what the user said, finds the matching gap items, locates the correct markdown files in the library, and rewrites them with the new information plus an updated verification date.
The same skill runs on the submissions folder when an approver triggers it — different source, same logic. The skill does not blindly paste; it reasons about what already exists, what is new, and what should be merged where.
Ingestion is reasoning, not copy-paste. The skill decides which entries to update, which gaps to close, and what to leave for human review.
Ingesting knowledge means dumping raw text into a file. The skill extracts what belongs in the KB and discards the rest — your meeting transcript becomes a structured policy entry, not a wall of text.
Why-Tracking — capturing rationale, not just facts
The third demo question — "what uptime do we require from critical vendors, and why is it set at that level?" — exposes the most powerful feature: the KB stores not just numbers, but reasoning. When the ingest skill processes a meeting transcript, it tries to extract the why behind decisions: what was discussed, what was rejected, what tradeoffs were accepted.
That is why the right input for ingestion is a meeting recording or email thread, not a polished policy doc. The richer the raw source, the more rationale the skill can capture — and that rationale becomes searchable knowledge.
A living KB stores decisions and their rationale, not just outcomes. The "why" is what makes the answers useful months after the meeting.
Feed the agent your finished policy documents and you get the best KB. The opposite is true — feed it raw transcripts and emails so it can capture the reasoning, not just the conclusion.
Flashcards — Applied
02 · Applied
AI Skills
tap to reveal →
Discrete agent-callable capabilities the KB exposes (e.g. log gaps, ingest knowledge, find experts). The agent picks which to invoke from a user question; users do not call them by name.
← tap to flip back
02 · Applied
Asking Questions
tap to reveal →
The entry point to the KB. The agent grounds answers on the document library, cites the source file, and can silently fire side-effect skills like gap logging.
← tap to flip back
02 · Applied
Knowledge Decay
tap to reveal →
Automatic freshness handling. Each file carries a last-verified date; the context file defines stale thresholds. Stale entries trigger verification tasks routed to owners.
← tap to flip back
02 · Applied
Ingest Knowledge Skill
tap to reveal →
The skill that promotes raw input into canonical entries. Reasons about what already exists, merges new info into the right files, updates verification dates, and resolves matching gaps.
← tap to flip back
02 · Applied
Why-Tracking
tap to reveal →
Ingestion captures the rationale behind a decision, not just the outcome. Best fed raw transcripts and email threads — the agent extracts both the number and the reasoning that produced it.