1. Where onboarding stands
Like a house that passed inspection, with a new, higher bar just announced.
- Completed all onboarding tasks you asked for. All of them are live.
- Two leftovers: Documents can't be downloaded yet, and the existing knowledge base isn't read.
- Fede's replay shows Clara can answer about 50% of Situs's past questions.
- Go-live bar he set: 80 to 90% answerable.
- The gap is depth of discovery. The data is in AppFolio and conversations.
Proves: the 2026-10-15 founders' sync transcript (#transcripts) and today's prod checks.
2. The idea in one picture
Like moving in: empty every drawer onto the floor, then sort into labeled boxes.
- Every PMS has the same things: messages, documents, guest cards, work orders.
- So one layer pulls raw data, and the rest doesn't care which source.
- A stronger agent turns the dump into one row per policy.
- The onboarding hub reads that sheet. Customers review only the last 10 to 20%.
- A human review step always stays in.
Proves: the 2026-10-15 sync, where Fede and Gera agreed on dump, then reason, then pre-fill.
3. Phase 1: pull everything
Like a mover who boxes everything first, then asks what to keep.
- AppFolio first: API reports, then the robot for screens the API lacks.
- Then guest cards, notes, maintenance, years of messages, lease documents.
- Email and Drive next, when the customer connects them.
- Each item keeps when it was seen, so newest wins later.
- Cadence is per source: one-time at onboarding, or weekly, or hourly.
Proves: today's AppFolio sync is already a connector; Fede has 3 years of Situs data in S3.
4. Phase 2: fill the policy sheet
Like a census form: every policy gets a line, even if it's blank.
- Rows: pet fee, application fee, deposit, renewal terms, office hours, and so on.
- Each row: value, detail, source, seen-at date, agreement count, conflicts.
- Conflicts (staff quoting $28, $35, $40) go to the customer to pick.
- Export it as a sheet you can open, plus the onboarding pre-fill.
- Coverage % tells us how far we got.
Proves: the dry run already prints value, source and evidence for 8 questions; this widens it.
5. Phase 3: dig deeper until the replay passes
Like studying with past exams until you score high enough.
- The real "done" signal is the replay score, not how full the sheet looks.
- Each miss points at a gap: a missing source or a shallow read.
- Overnight agents can work the gap list.
- The PMS knowledge base repo tells them where things live in AppFolio.
- Stop when the replay clears 80 to 90%.
Proves: Fede's replay harness (he's pushing it to the skills repo) and the pms-knowledge-base repo.
6. How we get there
Like building a road one stretch at a time, driving it as we go.
- Step 1: get Fede's replay harness, reproduce Situs's 50%, list the gaps.
- Step 2: close Situs's gaps by hand with Claude, writing down each path.
- Step 3: turn those paths into connectors in the new repo.
- Step 4: widen the dry run into the full policy sheet.
- Step 5: write the ADR before the repo, as our rules require.
Proves: the order Fede described: by hand first, then automate what you learned.
7. What to call it
Like naming a filing cabinet: it should say what goes inside.
- dossier (lean): the complete file we build on a company. Connectors fill it, the replay grades it.
- intake: plain and obvious. It's how a new customer's information comes in.
- fieldbook: a field researcher's notebook. Fits the policy sheet idea.
- census: counts every policy, blank or not.
- harvest: gathers everything in. Friendly, but vaguer.
Proves: no existing system by these names in SYSTEMS.md (checked 2026-10-15).
A1 · Appendix: what exists today vs what's new
| Layer | Exists today | New |
|---|---|---|
| Connectors (phase 1) | AppFolio sync (a connector), browser agent (a way to reach screens), Fede's S3 dumps (3 years of guest cards, maintenance, messages), pms-knowledge-base (how AppFolio works) | One interface for every source; Gmail, Outlook, Drive, Monday connectors; per-source schedule (Temporal) |
| Policy sheet (phase 2) | onboarding:prefill-dry-run and the onboarding-backfill skill: value, source, evidence for 8 questions | Every policy; sheet export; coverage % |
| Reasoner | "Most common amount" logic for fees; Hand off to client for the hard ones | An agent pass over the raw dump; freshest wins; conflicts listed |
| Pre-fill | Saved answer → pulled value → Propflow default, with source labels (live) | Nothing new |
| Gate | Fede's replay harness (local, being shared) | Run it as a standard step before go-live |
Where it lives
A separate repo, as agreed in the sync: propflow-ingest (private; created 2026-10-15, decision record ADR-0059). The browser-agent repo stays the transport. Onboarding and the replay sit on top. The AppFolio page "atlas" stays parked: the sheet here maps policies, not screens.
Rules this follows
- An ADR before the repo (cross-system decision).
- A SYSTEMS.md entry when it ships.
- Never decrypt a customer's stored login; use the API and normal clients.
- Nothing saves to a customer's settings without a person's click.