Add Paste MCP ambient layer design spec
Four-phase architecture: Content Outbox + Call Prep (Phase 1, immediate), Brand Vault (Phase 2), Session Handoff (Phase 3), Agent Trays (Phase 4). Phase 1 starts now; Phase 3 gated on auth stability across 5+ sessions. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
parent
755d43d5c9
commit
43dcab2f68
1 changed files with 141 additions and 0 deletions
|
|
@ -0,0 +1,141 @@
|
|||
# Paste MCP Ambient Layer — Design Spec
|
||||
**Date:** 2026-06-08
|
||||
**Scope:** STTIL Solutions infrastructure (all projects)
|
||||
**Status:** Approved for Phase 1 implementation
|
||||
|
||||
---
|
||||
|
||||
## Problem
|
||||
|
||||
Every knowledge system in the STTIL stack (TriLane, MemPalace, Obsidian, NocoDB) requires a query. Content produced by Claude lives only in the chat window until manually copied. Call briefs generated by pilot-prep disappear after the session. There is no ambient layer — one that is always live, always accessible in every app, zero friction to use.
|
||||
|
||||
Paste MCP fills that gap. It is the bridge between what AI produces and where it actually gets used.
|
||||
|
||||
---
|
||||
|
||||
## Design Principle
|
||||
|
||||
Paste is always **downstream** of authoritative systems, never upstream.
|
||||
|
||||
- MemPalace KG: authoritative for brand facts
|
||||
- NocoDB: authoritative for leads and content status
|
||||
- TriLane: authoritative for deep research retrieval
|
||||
- Paste: working copy of what is in active use right now
|
||||
|
||||
Paste holds the hot layer. It does not replace any deep storage system.
|
||||
|
||||
---
|
||||
|
||||
## Phased Architecture
|
||||
|
||||
### Phase 1: Content Outbox + Call Prep
|
||||
**Trigger:** Implement now. No startup dependency, no auth risk.
|
||||
|
||||
**Pinboards created:**
|
||||
- `Content Outbox` — staging layer for all publishable content
|
||||
- `Call Prep` — pre-call brief for every pilot and investor call
|
||||
|
||||
**Content Outbox — how it works:**
|
||||
1. linkedin-post skill finishes drafting → pushes draft to Content Outbox via `mcp__paste__create_item`
|
||||
2. linkedin-carousel skill finishes drafting → same
|
||||
3. Kisa reviews draft in Paste app, pastes directly to LinkedIn or wherever it's going
|
||||
4. Claude updates NocoDB status to Published in the same move
|
||||
5. Item deleted from Content Outbox — the pinboard stays clean, not an archive
|
||||
|
||||
**Call Prep — how it works:**
|
||||
1. pilot-prep skill finishes → pushes call brief to Call Prep pinboard
|
||||
2. Brief is in clipboard history before the call starts, accessible from any app or device
|
||||
3. After the call, Granola pipeline runs as normal (unchanged)
|
||||
4. Call Prep item cleared after the call
|
||||
|
||||
**Skills updated in Phase 1:**
|
||||
- `linkedin-post`: add final step → `mcp__paste__create_item` to Content Outbox
|
||||
- `linkedin-carousel`: add final step → `mcp__paste__create_item` to Content Outbox
|
||||
- `pilot-prep`: add final step → `mcp__paste__create_item` to Call Prep
|
||||
|
||||
**Granola note:** Granola ingestion (post-call) is unaffected. The pipeline (Obsidian sync → TriLane → NavBot) stays exactly as is. Call Prep is pre-call only.
|
||||
|
||||
---
|
||||
|
||||
### Phase 2: Brand Vault
|
||||
**Trigger:** After Phase 1 proves stable across multiple sessions.
|
||||
|
||||
**Pinboard created:** `Brand Vault`
|
||||
|
||||
**Contents (~12-15 items, all KG-verified):**
|
||||
- CERT 2019 framing: "nearly 1 in 3" (32.8%), "over two-thirds from docs" (68.6%)
|
||||
- CGS prepay stat: ~30.86% error rate, Q2 2024
|
||||
- Badge labels: At Risk (SUPPLY_LAPSED/VISIT_REQUIRED/RENEWAL_CRITICAL), Action Needed (RENEWAL_ELEVATED/RENEWAL_SOON/TRANSFER_PENDING), Clear to Ship (RESUPPLY_READY), On Track (ACTIVE)
|
||||
- Supply continuity naming: Supply Lapsed / Renewal Due / Resupply Ready / Active / Outreach Worklist
|
||||
- Renewal Due cadence: 60d patient, 45d prescriber
|
||||
- Pricing tiers: Worklist $149-199, Connected $349-499, Intelligence $699-999+usage
|
||||
- Pitch anchor: "$400-500M/yr spent documenting the problem Signal solves; Worklist costs $1,788/yr"
|
||||
- CMS PA exemption: ≥90% affirmation rate = PA exemption, June 1 2026 first cycle
|
||||
- Brand sign-off block: "Signal. / Powered by STTIL Solutions"
|
||||
|
||||
**How it stays current:**
|
||||
The Decision Protocol (CLAUDE.md) already triggers `kg_add` when a brand fact is confirmed. Phase 2 adds one step: after `kg_add`, update or create the corresponding Vault item in Paste. KG remains authoritative — Vault is the working copy.
|
||||
|
||||
**What it replaces:** The need to query me for locked stats and copy blocks during active work.
|
||||
|
||||
---
|
||||
|
||||
### Phase 3: Session Handoff
|
||||
**Trigger:** After Paste auth proves reliable across 5+ consecutive sessions.
|
||||
|
||||
**Pinboard created:** `Hot Context`
|
||||
|
||||
**What it contains:** Curated ~200-word summary of what is in active motion — not history, not everything. Current build status, who is being called this week, what is blocked, immediate next actions.
|
||||
|
||||
**navaigate-session-brain update:** Add final step — push Hot Context summary to Paste.
|
||||
|
||||
**Session startup update (CLAUDE.md):** Add Paste Hot Context read at the top of the startup protocol. If Hot Context exists, skip `project_next_steps.md` and `project_vc_traction.md` reads — those become fallbacks for auth failures or empty pinboard. KG queries and NocoDB live pulls stay unchanged.
|
||||
|
||||
**Fallback rule:** If Paste auth fails or Hot Context is empty, the startup protocol falls through to the existing file reads with no change in behavior. Paste is never in the critical path until auth is proven.
|
||||
|
||||
**Token savings:** ~400-500 tokens per focused session. Startup speed improvement on days with a clear working agenda.
|
||||
|
||||
---
|
||||
|
||||
### Phase 4: Agent Trays
|
||||
**Trigger:** After AiOS restructure completes and division agents are running.
|
||||
|
||||
**Pinboards created:** `Signal Tray`, `Content Tray`, `STTIL Tray`, `Infrastructure Tray`
|
||||
|
||||
**How it works:** Each AiOS division agent writes its current outputs and status to its tray. The morning skill reads all four trays as part of the briefing — two-sentence status per division, without triggering full agent runs.
|
||||
|
||||
**Scaling:** Med Tracker, UMSIS, and Game Center each get a tray pinboard when they activate. No redesign needed — add pinboard, wire agent, done.
|
||||
|
||||
---
|
||||
|
||||
## Full Pinboard Map
|
||||
|
||||
| Pinboard | Phase | Written by | Read by |
|
||||
|----------|-------|-----------|---------|
|
||||
| Content Outbox | 1 | linkedin-post, linkedin-carousel | Kisa (paste to publish) |
|
||||
| Call Prep | 1 | pilot-prep | Kisa (during calls) |
|
||||
| Brand Vault | 2 | KG update flow | Kisa, copy-review |
|
||||
| Hot Context | 3 | navaigate-session-brain | Session startup protocol |
|
||||
| Signal Tray | 4 | Signal agent | morning skill |
|
||||
| Content Tray | 4 | Content agent | morning skill |
|
||||
| STTIL Tray | 4 | STTIL agent | morning skill |
|
||||
| Infrastructure Tray | 4 | Infrastructure agent | morning skill |
|
||||
|
||||
---
|
||||
|
||||
## What This Does NOT Change
|
||||
|
||||
- MemPalace KG: still authoritative for brand facts
|
||||
- TriLane: still handles deep research retrieval and session logs
|
||||
- NocoDB: still authoritative for leads and content calendar
|
||||
- Granola pipeline: unchanged
|
||||
- Memory .md files: permanent sections (preferences, feedback, paths) stay as-is; rolling sections (next steps, VC status) become fallbacks in Phase 3
|
||||
|
||||
---
|
||||
|
||||
## Pitfalls to Watch
|
||||
|
||||
1. **Auth reliability** — Paste auth does not yet persist reliably across sessions. Phase 3 is blocked until this proves stable. Phase 1 and 2 have no startup dependency and are safe to build now.
|
||||
2. **Dual maintenance** — Brand Vault items must be written FROM KG, never independently. Any fact in Paste must trace back to a KG entry.
|
||||
3. **Pinboard sprawl** — Keep pinboards small and curated. Paste is not a search system. If any pinboard grows past 20-25 items, it needs pruning or a sub-pinboard.
|
||||
4. **Outbox hygiene** — Content Outbox must be cleared after each item is published. Stale drafts in the Outbox create confusion.
|
||||
Loading…
Reference in a new issue