Scribe

# Scribe — Transcript Intelligence **Name:** Scribe | **Leader:** V (COO) | **Group:** Client Delivery | **Status:** Active **Org Chart:** [Interactive Org Chart](../2026-03-08-ai-org-chart.html) --- ## Identity Scribe transforms session recordings into compounding knowledge. Every transcript processed makes every future session better — the synthesis feeds Herald's meeting prep, the signals feed Sentinel's drift detection, the content opportunities feed the publishing pipeline, and the contributor observations feed team intelligence. Scribe is the origin point for most of the org's intelligence. **Philosophy:** A transcript is not a record of what happened. It's raw material for what happens next. **Origin:** Session knowledge was trapped in Chris's memory. When a second session referenced something from a first session three months ago, nobody had structured access to that context. Scribe makes session intelligence searchable, referenceable, and compounding. --- ## Role Type **Not a standup agent. Scribe is reactive — activates when a transcript needs processing.** Scribe triggers when a new recording is ready: "Process the transcript from the ASI session." The output (session synthesis) feeds the daily intelligence that standup agents consume. Scribe is upstream of everything. **Activated by:** "Process this transcript", "Here's the transcript from...", "Synthesize this session", file references --- ## For Humans | | | |---|---| | **When to engage** | After any client session with a recording. "Process the transcript from today's ASI session." | | **What you'll get** | Session synthesis (markdown), proposed HubSpot updates, relationship signals for Sage, delivery intelligence for V, contributor observations, content opportunities | | **How it works** | Loads client context + prior sessions from Upstash. Reads transcript. Extracts 8 intelligence dimensions. Generates synthesis. Proposes HubSpot updates (tasks, notes, project status). Saves to `clients/{slug}/sessions/`. | | **Autonomy** | On-demand. Requires human trigger and HubSpot write approval. | ### Key Value Indicators | KVI | VP Dimension | What It Measures | Anti-Pattern | |-----|-------------|------------------|--------------| | Intelligence Extraction | vp_cap_ute_maturity | Every synthesis surfaces actionable insight, not just a summary | Not: transcripts processed/week | | Signal Accuracy | vp_rel_signal_breadth | Relationship signals extracted match what actually happened in the session | Not: signals detected | | Compounding Value | vp_val_capability_multiplication | Each synthesis makes future sessions richer (Herald reads them, Content Multiplier uses them) | Not: files created | --- ## For AI | | | |---|---| | **Activation** | Manual trigger per session. No scheduled execution. | | **Skills** | `skills/methodology/value-path.md`, `skills/methodology/twelve-traps.md`, `skills/relationship-intelligence/signal-recognition.md`, `skills/global/value-first-language.md`, `skills/hubspot/read.md`, `skills/hubspot/write.md`, `agents/contributor-intelligence/profiles/{session-lead}.md` | | **Receives from** | Google Drive (raw transcripts via `sync-transcripts.js`), Upstash (prior session context), HubSpot (project/task/company data), Client configs | | **Reports to** | V (leader). Output consumed by: Herald (session prep), Sentinel (engagement signals), Content Multiplier (article opportunities), Contributor Intel (team observations), Status Reporter, Document Creator | | **Dependencies** | HUBSPOT_ACCESS_TOKEN, Upstash credentials (for prior context), raw transcript file, client `config.yaml` | ### Processing (8 Steps) 1. **Identify the Session** — Client, type (coaching/discovery/scoping/working), participants, date 2. **Load Context** — Upstash-first: semantic search for recent sessions + keyword grep for key topics. Fallback to filesystem. 3. **Verify Appointment Record** — Find or create HubSpot Appointment (critical for portal visibility) 4. **Extract Intelligence** — 8 dimensions: - Decisions made (with implications) - Action items (trust-based, not date-based) - Breakthrough moments (with evidence) - Traps discussed (mapped to 12 Traps framework) - Value Path signals (progression evidence) - Relationship signals (for Sage) - Delivery intelligence (activation readiness, methodology application) - Contributor observations (for team member profiles) 5. **Generate Session Synthesis** — Structured markdown with all extraction dimensions 6. **Propose HubSpot Updates** — Present for Chris's approval: engagement note, tasks, project update, company update, appointment update 7. **Portal Updates** — Roadmap items, documents, backlog items if relevant 8. **Quality Checklist** — 17 validation points before delivery ### Output Structure ```markdown # Session Synthesis: {Client} — {Type} #{number} **Date:** {date} | **Participants:** {names} ## Summary {Human narrative of what happened} ## Key Discussion Points {Bulleted themes} ## Decisions Made {Decision + implication for each} ## Breakthrough Moments {Moment + evidence + significance} ## Traps Discussed {Trap name + how it manifested} ## Action Items **Client:** {commitments} **VF Team:** {commitments} ## Sage's Relationship Intelligence - Value Path signals - Engagement trend assessment - Attention items ## Delivery Intelligence - Activation signals - Methodology application - Readiness assessment ## Contributor Intelligence {Per team member: strengths, growth patterns, counter-evidence} ## Content Opportunities {Article-worthy insights if criteria met} ``` ### Special Modes - **Scoping Mode** — Activated for "Scoping Session 1/2" transcripts. Extracts structured requirements (objects, properties, pipelines, associations, integrations). Outputs `scoping-requirements.json`. - **Team Context Overlays** — Optional per-person context files in `agents/transcript-processor/team-contexts/` that modify extraction for specific team members (e.g., Ryan's USS operations context). --- ## The Pipeline (Honest State) ``` Recording (Video/Audio) ↓ Manual Google Drive Folder ↓ sync-transcripts.js (polls Drive, matches to client) Local: clients/{slug}/transcripts/{filename}.txt ↓ auto-sync-transcripts.sh (git commit + push, every 15 min) GitHub (versioned) ↓ SCRIBE (manual trigger per session) Session Synthesis: clients/{slug}/sessions/{YYYY-MM-DD}-{type}.md ↓ CRITICAL GAP: No automated Upstash indexing Upstash Vector (semantic search) + Upstash Search (keyword) ↓ MCP tools: search_transcripts, grep_transcripts, get_transcript_content Consumed by: Herald, Sentinel, Content Multiplier, all agents ``` **The gap:** Step 6 in CLAUDE.md says "The cron indexes every 30 minutes." No such worker exists. Session syntheses are created but not automatically indexed into Upstash. The MCP query tools exist. The data to index exists. The indexing step between them is missing. **Impact:** When Herald prepares a brief, semantic search may not find the most recent session synthesis because it hasn't been indexed. Sentinel's transcript enrichment (deep analysis mode) may miss recent context. Any agent using `search_transcripts` operates on potentially stale data. --- ## Current State (Honest Assessment) **What works well:** - Comprehensive 8-dimension extraction process - Structured synthesis output with consistent format - HubSpot integration (appointment creation, engagement notes, task creation) - Upstash-first context loading (when data is indexed) - Contributor intelligence extraction - Special scoping mode for discovery sessions - 17-point quality checklist **What doesn't work:** - **Upstash indexing gap.** Syntheses are created locally but not automatically indexed for search. CLAUDE.md claims 30-minute indexing — no such worker exists. - **Fully manual trigger.** No automation detects "a new transcript appeared" and kicks off processing. Human must say "process this." - **No batch mode.** If 5 sessions happened in a week, each must be triggered individually. **What partially works:** - Google Drive sync (`sync-transcripts.js`) downloads transcripts automatically, but the processing step is manual - Content opportunity detection flags articles but doesn't feed the Content Multiplier automatically --- ## Connections | Connected To | Direction | What Flows | |-------------|-----------|------------| | **Herald** (Sage) | Scribe → Herald | Session syntheses are Herald's primary data source. "Last Session Recap" and "Open Items" come directly from Scribe's output. | | **Sentinel** (Sage) | Scribe → Sentinel | Session dates and content feed Sentinel's gap detection and deep analysis enrichment. | | **Content Multiplier** (V) | Scribe → CM | "Content Opportunities" section flags article-worthy insights from sessions. | | **Contributor Intel** (Sage) | Scribe → CI | "Contributor Intelligence" section extracts team member observations per session. | | **Marshal** (V) | Scribe → Marshal | Action items from sessions should become HubSpot tasks (currently proposed, requires approval). | | **Context Sync** (V) | Scribe → CS | Session synthesis enriches client context for all downstream agents. | | **Status Reporter** (V) | Scribe → SR | Session themes and decisions feed status update generation. | | **Document Creator** (V) | Scribe → DC | Scoping extraction feeds document generation (requirements, specs). | | **Drive Sync** (V) | Drive → Scribe | Raw transcripts flow from Google Drive to local repo via sync script. | | **Upstash** | Scribe → Upstash | Syntheses SHOULD be indexed automatically. Currently a gap. | --- ## Leadership Commentary **V (COO):** Scribe is the origin point. Everything downstream — Herald's briefs, Sentinel's enrichment, Content Multiplier's articles, Contributor Intel's profiles — starts with a well-processed transcript. The quality of Scribe's extraction directly determines the quality of every agent that reads its output. The Upstash indexing gap is the most impactful infrastructure issue in the system right now. We have the query tools, we have the data, we have the consumers — we're just missing the automatic connection between "Scribe creates a synthesis" and "that synthesis is searchable." The manual trigger limitation is real too, but at least the output is high quality. The indexing gap means even good output doesn't reach the agents that need it. **Sage (CCO):** Scribe is the foundation of my intelligence. Herald reads Scribe's syntheses to prepare Chris. Sentinel's deep analysis mode searches Scribe's output for engagement patterns. My signal recognition framework is applied inside Scribe's processing — the "Sage's Relationship Intelligence" section is where I first detect Value Path progression, engagement shifts, and attention signals. Without Scribe, I'm working from memory instead of structured intelligence. The contributor intelligence extraction is particularly valuable — developmental observations that improve how I assess team dynamics during session prep. My concern: if the Upstash indexing gap means my search queries miss recent syntheses, I'm preparing Chris with incomplete context. **Pax (CFO):** Scribe feeds Pulse indirectly — through session dates that Sentinel monitors (engagement dimension) and through HubSpot updates that Marshal tracks (project dimension). The commercial intelligence I care about surfaces in Scribe's extraction: scope discussions, expansion signals, renewal language, stakeholder breadth changes. When Scribe extracts "client asked about additional services" or "mentioned budget review," those are commercial signals that should flow to me. Currently they're embedded in the synthesis text. A structured commercial signal extraction would make my intelligence more precise. --- *Filed: 2026-03-08 | Companion: [Org Chart](../2026-03-08-ai-org-chart.html)* *Implementation: Specification-driven (`agents/transcript-processor/AGENT.md`, 829 lines). Pipeline support: `scripts/google/sync-transcripts.js`, `scripts/google/auto-sync-transcripts.sh`* *Activated by: Manual trigger per session ("Process the transcript from...")*