Diana Snodgrass — Institutional Knowledge Transfer & AI Second Brain
Analysis prepared by Michelle | April 8, 2026 | Confidential — CAA Internal
The Short Version: This is absolutely doable, and it's exactly the right time to start. Diana's knowledge falls into two buckets: the stuff that can be captured (processes, rules, precedents, vendor contacts, decision logic) and the stuff that can't (judgment calls built from 20 years of reading a room, relationships, and institutional intuition). A well-designed second brain can handle 70–80% of what Diana does day-to-day. The other 20% is why you keep her part-time in 2027.
1. What CAN Be Captured vs. What Can't
Good candidates for capture
- Step-by-step processes: open enrollment, 401k administration, COBRA, onboarding/offboarding, compliance filings
- Vendor and contact directories: insurance brokers, legal counsel, 401k administrator, state regulators
- Policy interpretations: "what does our handbook say about X, and why did we decide that"
- Regulatory calendar: filing deadlines, renewal windows, required notices by law
- Decision trees: when to call the lawyer vs. handle in-house, when a situation is a PIP vs. termination
- Templates and precedents: warning letters, accommodation requests, leave approvals, contract language
- FAQ bank: the 50 questions that come up over and over
- History and context: past legal issues, settlements, plan changes, why certain policies exist
Hard to capture
- Judgment built from experience ("this feels like it's going somewhere bad — call legal now")
- Relationship nuance with specific employees, vendors, or regulators
- Reading tone in a legal situation or negotiation
- Knowing when the written policy isn't actually how you handle something
- Anything requiring human empathy — an employee disclosing a serious illness, a difficult termination
Bottom line: processes, rules, and reference material transfer well. Wisdom and instinct don't fully transfer anywhere — that's what part-time coverage in 2027 is for.
2. How Knowledge Capture Works — Self-Directed Sessions
Rather than structured facilitator-led meetings, Diana works through prepared prompt questions on her own schedule. The goal is roughly three hours per week — but that doesn't need to be three continuous hours. It can be 15 minutes in the morning, an hour at lunch, a half-hour at the end of the day. The work is cumulative.
How it works in practice: Each domain comes with a prepared set of prompt questions. Diana answers them in whatever format is most natural to her — typed notes, voice memo, even a quick email. The answers get captured, organized, and eventually loaded into the knowledge base. Her job is to answer. Someone else's job is to organize and maintain what she produces.
Recommended domains
- HR & Employment — hiring, discipline, termination, leaves, accommodations, handbook policy
- Health Insurance — plan details, open enrollment, COBRA, vendor contacts, employee FAQs
- 401k Administration — plan rules, compliance tests, vendor contacts, common employee questions
- Legal & Regulatory — when to escalate, regulatory contacts, filing calendar, past issues
- Compliance — required notices, reporting deadlines, state-specific rules
What each domain's prompt set covers
- Process flows — "Walk me through step by step what happens when…"
- Decision logic — "When you see X, how do you decide between Y and Z?"
- Reference material — vendors, contacts, account numbers, portal access
- Context and history — "Why do we do it this way? When did that change? What happened that time we got this wrong?"
- Edge cases and exceptions — "What are the situations that don't follow the normal process?"
Prompt questions are released one domain at a time — not all at once. Starting with the highest-risk domains (legal and compliance first, not the easiest ones) ensures that if the timeline compresses, the most important material is already captured.
3. The Avatar Concept
A static document library is useful. An AI agent that answers questions the way Diana would is transformative.
The goal is to build a Diana Second Brain Avatar — an AI agent loaded with everything captured through the prompt sessions. CAA staff who relied on Diana for guidance could ask the avatar directly: "What do I do when an employee requests FMLA for a mental health reason?" or "Who do I call if we have a workers' comp situation in Georgia?" and get answers grounded in Diana's actual knowledge and CAA's actual practices.
Why this beats a knowledge base: Nobody searches a knowledge base when they're in the middle of a situation. They ask a person. The avatar preserves that interaction model — it feels like consulting Diana, not searching a filing cabinet. That's the difference between a system people actually use and one that goes stale after six months.
What the avatar can do
- Answer procedural questions: "What's the process for putting someone on a PIP?"
- Surface the right contacts: "Who handles our 401k compliance testing, and what's their number?"
- Flag when to escalate: "This sounds like a situation where you'd want to loop in legal before responding"
- Explain why, not just what: "We handle it this way because of a situation we had in 2019 where…"
- Provide templates and starting points for common documents
What it won't replace
- Real-time judgment on novel or sensitive situations — those still need a human
- Empathy in difficult conversations with employees
- Legal advice — it can tell you when to call the lawyer, not what the lawyer would say
Architecture
The avatar runs on a RAG (retrieval-augmented generation) architecture — the same proven approach used for Professor Cody (building codes AI). Diana's captured knowledge lives in a vector database; the avatar retrieves relevant material and uses it to answer questions accurately rather than hallucinating. New material added over time automatically improves the avatar's responses.
Note: Naming the avatar "Diana" — or any variant using her name — should be confirmed with Diana directly before deployment. Most people feel honored by the idea of their knowledge living on this way, but it's worth a conversation first.
4. Ownership and Maintenance
The knowledge base needs a keeper — someone who reviews what Diana produces, organizes it, keeps it current as things change, and decides what goes into the avatar. Laws change, plans renew, policies evolve. A knowledge base that doesn't get maintained becomes unreliable fast.
Initially, Rusty serves as the keeper: reviewing captured material, making judgment calls about what belongs in the system, and maintaining quality control. Over time, as the system matures, that role can be delegated or distributed. But having clear ownership from day one is what separates a functioning system from an abandoned one.
5. Technology Options
| Option | Description | Best For | Complexity |
|---|
| A — Doc Library + AI Chat | Organized folders + Notion AI, M365 Copilot, or custom GPT with uploaded docs | Starting point; low cost; no coding | Low |
| B — Custom RAG (recommended) | Vector DB (ChromaDB) + Claude/GPT-4 retrieval layer — same architecture as Professor Cody | More control, better accuracy, supports the avatar concept fully | Medium |
| C — Searchable Wiki | Notion, Confluence, SharePoint — no AI layer | Fast to build; no AI value until later | Low |
Recommendation: Target Option B from the start — the avatar concept requires it, and CAA already has the infrastructure (ChromaDB, OpenClaw agents). Start building the document library now and stand up the RAG layer once enough content exists to make it useful. Don't let technology selection slow down the knowledge capture work — the two happen in parallel.
6. Realistic Limitations
The avatar will handle well
- "What's the deadline for open enrollment notices?"
- "What's our process when an employee requests FMLA?"
- "Who's our 401k contact and what's the plan number?"
- "What does our handbook say about remote work?"
- "Walk me through how to handle a written warning"
It will struggle with
- Novel situations outside documented precedent
- Anything requiring legal judgment in real-time (call a lawyer)
- Sensitive HR situations requiring human empathy and discretion
- Multi-factor judgment calls that depend on who the people involved are
- Anything that changes after the last knowledge update (this is why maintenance matters)
7. Recommended Phased Approach
Phase 1 Inventory & First Domain — April–June 2026
Map all of Diana's domains. Prioritize by risk. Release first prompt set (legal/compliance). Diana begins answering on her own schedule. Rusty reviews and organizes responses.
Phase 2 Deep Capture — July–October 2026
Release prompt sets domain by domain. Diana works through them at ~3 hrs/week. New questions added as gaps surface. Avatar architecture stands up once enough content exists.
Phase 3 Build & Test the Avatar — October–November 2026
Load captured knowledge into the RAG system. Test with real questions from staff. Gap-fill based on what it can't answer. Refine with Diana before she transitions.
Phase 4 Handoff & Validation — December 2026
Diana reviews the full system. Final gaps addressed. Avatar goes live for CAA staff. Diana moves to part-time, available for complex calls the avatar can't handle.
Phase 5 Part-Time Overlay — 2027
Diana available for judgment calls. Avatar handles routine. Her hours used to update and refine based on what's actually being asked.
8. Timeline Reality Check
Nine months (April through December) is workable but not comfortable. Realistically:
- ~3 hours/week of Diana's time, self-directed across her schedule
- Over 9 months = roughly 100 hours of knowledge capture
- Enough to cover 4–5 domains thoroughly if started in April or May
Where it goes wrong
- Summer schedule compressing the timeline if capture doesn't start soon
- Waiting to release prompt questions — Diana can't answer questions she hasn't received
- Underestimating editing/refinement time after Diana's responses come in
Where you can accelerate
- Release prompt questions in batches and let Diana get ahead of the schedule
- Use AI transcription if Diana prefers voice memos over typing
- 70% complete before she retires beats 100% complete that never gets finished
Next step: Release the first domain prompt set — Legal & Compliance — and let Diana begin. Everything else follows from there.
Charles Abbott Associates | Prepared by Michelle, April 8, 2026 | Confidential