SyF Second Brain — Build Strategy

Cold-start index pattern (NEXT → INDEX → Context)

Aria's three-layer cold start (NEXT.md → MEMORY.md → lane files) maps directly to SyF. SyF-NEXT is the single most important page — it's what makes every session feel like a continuation, not a restart. If Topher maintains nothing else, he maintains this. The rest is reference.

Paste-shape rule — self-completing prompts, no inline fill-in

Long prompts pasted into a chat UI often collapse to "[pasted N lines]" — you can't see or edit inline variables. Every SyF prompt must end with "confirm you've loaded the context, then ask me what we're working on today." The AI waits for the real ask rather than trying to pre-fill a template Topher forgot to update. This is why the Session Starter is designed as a loader, not a query.

Dashboard discipline — same-session updates, never stale

Every Aria session that closes a loop also updates the dashboard in the same commit. No exceptions — or urgency cards stay stale and trust erodes. For SyF: every SYF GPT session that produces an output (Synchrony Pay slide outline, Bijayta deck draft) must also update SyF-NEXT before Topher closes the browser. If the NEXT page is stale, the next session starts cold.

Tiered load depth — always Tier 0, branch into others

Aria never loads all 100KB of memory. She loads NEXT (always), then branches into relevant lane files. For SyF, SYF GPT has 128k context — enough for everything, but throwing everything in degrades response quality (LLM attention falls off in the middle of long inputs). Load SyF-NEXT + one Context page max per session. Add more only if the answer is clearly wrong.

Lazy delegation — route only when handoff cost < execution cost

From Linq's architecture: don't route work just because you can. For SyF: SYF GPT handles strategic synthesis and real-time problem-solving. Confluence handles audit trail and team-visible frameworks. SharePoint handles self-service. If a stakeholder needs to reference it in 30 days → Confluence/SharePoint. If it's tactical and one-time → SYF GPT only.

Building before verifying the signal

Aria built a zombie reaper that killed live sessions because the "idle" signal (transcript mtime) never updated — it reflected creation time, not activity. The M3 Ultra on-prem pilot was fully built before we confirmed the compliance story. For SyF: Before building toward ROVO automation, verify ROVO is enabled at Synchrony. Before designing around M365 Copilot, verify it's licensed. Don't build the Tier 3 Confluence structure assuming ROVO works — confirm first.

Claiming "done" without showing the work

Aria's feedback rule: never say you've done something without evidence visible to Topher. In SyF context: SYF GPT should never say "I've documented that meeting decision" without the doc being paste-ready in the response. Every deliverable must be in the output, not referenced as complete. This is doubly critical in a financial services environment where audit trails matter.

Stale context silently degrading answers

Aria surfaced stale urgency cards multiple times — items were resolved in memory but the dashboard still showed them as open. For SyF: every OneNote Context page has a "Last Updated: [date]" stamp at the top. SYF GPT should be instructed (in the Session Starter) to note if any loaded page is >30 days old and ask if it's still current. Don't let February's Synchrony Pay outline drive April's exec conversation.

Over-engineering before production use

The most dangerous trap. It's tempting to design the perfect Tier 3 Confluence structure, ROVO agent configuration, and SharePoint taxonomy before any of it has been tested against real work. Ship Tier 2 first. Use it for one real Bijayta deck and one real Synchrony Pay session. What's missing will be obvious. What seemed necessary will prove unnecessary. Build from evidence, not architecture diagrams.