meeting

Analyze meetings from Granola. Adaptive approach — the agent reads the material, asks what matters, and decides how to analyze.

When to invoke

User says: "/meeting" (list unprocessed meetings), "/meeting sync" (process all), "/meeting [search]" (find and process a specific meeting), "/meeting backfill" (re-process historical meetings with richer extraction).

Arguments: $ARGUMENTS (Optional: "sync" for batch mode, "backfill" to re-process historical meetings, or search term to find a specific meeting)

Usage

• /meeting — Interactive: list recent unprocessed meetings, pick one
• /meeting sync — Batch: process all unprocessed meetings
• /meeting [search] — Find and process a specific meeting by title
• /meeting backfill — Re-process already-ingested meetings with richer extraction

Orientation

Three principles, not fixed constraints:

1. The user knows what matters — ask before analyzing. "Just give me the action items" and "help me understand the political dynamics" should produce fundamentally different analyses.
2. The material tells you how to read it — let inputs shape approach. A 10-minute standup and a 90-minute strategy session call for different depth.
3. The output schema is the only constraint — everything else is yours to decide. The graph needs structure to index against. Everything between "load context" and "write files" is the free zone.

The output schema exists because the graph needs structure. Everything between "load context" and "write files" is yours.

You are the analyst. The meeting is your material. The user is your client. The graph is your institutional memory. Do good work.

Intent Harvesting

Before you touch the transcript, ask the user what they need.

Use 2-3 quick AskUserQuestion prompts, adapted to context — not a fixed questionnaire. Examples:

• What matters most from this meeting?
• Full depth or quick extraction?
• Connect to prior threads or fresh analysis?
• Anything I should know going in?

The intent shapes everything: which lenses to apply, how deep to go, what output sections to emphasize.

• "just give me the action items" → lightweight pass, actions-focused, skip dynamics/continuity
• "I need to understand the political dynamics" → deep dynamics lens, speaker-by-speaker conviction mapping
• "what changed from last time?" → continuity-heavy, evolution focus
• "full analysis" or no strong preference → balanced multi-lens approach

Skip conditions — don't harvest intent when:

• User already gave rich context with their invocation (e.g., /meeting sync or /meeting [search term] just action items)
• Batch/sync mode — use adaptive defaults (balanced approach)
• Backfill mode — historical, no user to ask

Assessment Lenses

A library of ways to read, not a checklist to complete. Reach for whichever lenses the material and user intent call for.

Substance (what was said)

Priorities, dependencies, events, evidence-backed positions, tradeoffs, open questions, confidence signals.

Classification vocabulary:

Category	Signals	Example
decision	"we decided", "let's go with", explicit choices, "the plan is"	"Use stdio for MCP transport"
finding	"turns out", "we discovered", "interesting that", realizations	"Neo4j HTTP is faster than Bolt"
pattern	"every time we", "I keep seeing", recurring themes	"All pricing discussions converge on usage-based gating"
action	"I'll do", "[name] will", "next step is", "TODO"	"Oz will write the MCP spec by Friday"
unknown	Referenced but not elaborated — gaps to fill	"Pricing discussion"

Key questions: What was actually decided? What depends on what? What evidence supports it?

Dynamics (how it was said)

Tone arc, conviction strength, alignment/divergence, interpersonal patterns, energy shifts.

Key questions: Where did energy spike or drop? What was stated with conviction vs explored tentatively? Who drove and who followed?

Continuity (where it fits)

Decision evolution, topic recurrence, open threads addressed or dropped, meta-patterns, organizational arc.

Key questions: What shifted from previous meetings? What keeps recurring? What was dropped?

Criticality (what's at stake)

Tensions between stated positions and revealed behavior. Contradictions within a single speaker's framing. Risks acknowledged vs risks visible but unspoken. The gap between what people say they'll do and the energy they bring to it.

Key questions: Where does stated confidence diverge from actual tone? What risks are visible but unnamed? What contradictions did nobody call out?

---

These are ways of reading, not agents to spawn or boxes to fill. A quick action-item extraction might only need Substance. A politically complex meeting might need all four. The user's intent and the material itself tell you which lenses matter.

Analysis Toolkit

What's available (tools, not steps):

• Inline analysis: Read all inputs, produce output directly. Cheapest, fastest.
• Scaffold-first: Panel → lightweight scaffold → guided transcript reading. Good middle ground.
• Parallel sub-agents: Spawn via Task tool for dimension-specific deep analysis. Use for long/complex meetings.
• Selective lenses: Skip what doesn't apply. A standup doesn't need Criticality.
• Hybrid: Mix approaches freely. Scaffold inline, then spawn one agent for the lens that needs depth.

Factors to consider (guidance, not rules):

Factor	Implication
Transcript length (>30K chars)	Sub-agents may help for long transcripts
Transcript length (<10K chars)	Inline is probably sufficient
Graph context density	Continuity lens value scales with context
Meeting complexity (attendees, threads)	More complex → more analysis depth
User intent (quick vs deep)	May collapse toward inline
Cost sensitivity	Budget-aware → inline

Output Schema

The only true constraint. The graph needs structure.

Enriched Artifact Schema

Each extracted artifact gets these fields:

Field	Source	Description
category	Analysis	decision / finding / pattern / action
title	Analysis	Short descriptive title
content	Analysis	Full description with substance
context	Analysis	What conversation led to this
rationale	Analysis	The "why" behind it
tradeoffs	Analysis	What was considered and rejected
confidence	Analysis	0.0-1.0 based on evidence strength
speaker	Analysis	"me" (microphone) vs "them" (system)
panel_corroborated	Analysis	Whether panel also captured this
topics	Analysis	2-5 tags for quest linking
open_questions	Analysis	Raised but unresolved
evidence_quote	Analysis	Supporting quote (max 120 chars)
urgency	Analysis	high / medium / low / null
importance	Analysis	high / medium / low / null
conviction_strength	Analysis	assertion / hypothesis / exploration / null
conviction_challenged	Analysis	Was this challenged in discussion?
evolution_type	Analysis	new / shifted / reinforced / reversed / null
evolution_context	Analysis	What it supersedes/reinforces
related_extracts	Analysis	Cross-references between items [{id, relationship}]

Confidence calibration: strong_agreement=0.9, data_backed=0.9, single_speaker=0.7, exploratory=0.5, contentious=0.6.

Meeting Intelligence Briefing

# Meeting Intelligence: {Title}

**Date**: YYYY-MM-DD
**Attendees**: {names}
**Source**: Granola ({doc-id})
**Approach**: {description of analysis approach chosen}
**Intent**: {what the user asked for}
**Tone**: {tone description} | Alignment: {score}

## Meta-Analysis

{3-5 paragraphs of opinionated analysis. Structure around four lenses:}

{1. **Heart of the conversation.** Not the agenda — the gravitational center. What were people circling, building toward, or working through? What question was actually being answered, even if nobody framed it that way? First paragraph, get there fast.}

{2. **What's new.** What emerged that wasn't obvious going in? A shift in thinking, a constraint that surfaced, a convergence nobody declared. If nothing genuinely new emerged, say so — "execution meeting, no new ground" is a valid and useful finding. Don't manufacture novelty.}

{3. **Actuality frame.** Given what we know about the team's current state — active quests, recent decisions, open threads from previous meetings — where does this meeting land? What does it change, accelerate, or block? Connect to live organizational state. Reference specific quests, prior decisions, or open threads by name when relevant.}

{4. **Recommended considerations.** Sharp, specific. Not "continue exploring X" — what concretely should happen, who should act, and why it matters now. Distinguish between urgent (blocks other work) and important (shapes direction).}

## Tone & Energy

{Only if Dynamics lens was applied}
- Overall: {tone description}
- Arc: {toneArc}
- Key moments:
  - {moment}: "{quote}" — {speaker}

## Priorities

| Item | Urgency | Importance | Owner | Evidence |
|------|---------|------------|-------|----------|
| ... | high | high | them | "quote" |

## Dependencies

- {blocker} → {blocked} ({owner})

## Dynamics

{Only if Dynamics lens was applied}
- Pattern: {pattern description}
- Alignment: {score}
- Tensions: {if any}

## Convictions

{Only if Dynamics lens was applied}
| Statement | Speaker | Strength | Evidence |
|-----------|---------|----------|----------|
| ... | me | assertion | "quote" |

## Decision Evolution

{Only if Continuity lens was applied and cross-meeting data exists}
- **{topic}**: {previous} → {current} ({trajectory})

## Cross-Meeting Patterns

{Only if Continuity lens found patterns}
- {topic}: discussed {N}x, {trajectory description}

## Analytical Tensions

{Places where different lenses point in different directions, or self-contradictions within a single speaker's framing. Not "cross-agent disagreement" — real tensions:}
{- stated confidence vs actual energy}
{- content assertions vs organizational history}
{- risks acknowledged vs risks visible but unspoken}
{Example: "Charlie advocates empowerment philosophy but the Windows accessibility gap means users can't actually act on it"}

{If no tensions exist, omit this section.}

## Actions

{Extracted, not turned into Artifact files}
| Owner | Description | Evidence |
|-------|-------------|----------|
| ... | ... | "quote" |

## Artifacts Extracted

{List of artifacts with confidence + quest links}

## Open Threads

{Unresolved items carried forward}

Register: A trusted colleague who attended the meeting and is briefing someone who wasn't there. Direct, opinionated, doesn't waste words on obvious things. Prioritizes insight over completeness.

Anti-patterns:

• "The meeting covered several important topics..." → you have nothing to say
• Restating what's already in the artifact list → the list exists for that
• Equal weight to everything discussed → prioritize ruthlessly
• "It would be good to follow up on..." → who, what, by when, or don't say it
• Ignoring graph context → the actuality frame is what distinguishes this from any summary tool
• Burying the lead → heart of the conversation goes first, not last

Actions

Extracted, not turned into Artifact files. Schema per action:

• owner: who is responsible
• description: what they committed to
• evidence_quote: supporting quote (max 120 chars)

Quality Standards

• No trivia extraction — only knowledge worth preserving
• Panel corroboration → panel_corroborated: true, transcript-only → false + confidence ≤ 0.7
• Evidence quotes: max 120 chars, prioritize signal over length
• Speaker attribution: microphone = "me", system = "them"
• Confidence calibration: strong_agreement=0.9, data_backed=0.9, single_speaker=0.7, exploratory=0.5, contentious=0.6
• Don't extract small talk, logistics, or trivia
• If a scaffold item has no transcript discussion at all, omit it (don't fabricate)

Infrastructure

Mechanical steps that stay precise. Exact Cypher queries, exact file paths, exact state management.

Canonical ingest projection (mandatory)

Write the approved meeting and extracted artifacts to one durable egregore-knowledge-projection/v1 manifest under memory/ingest/knowledge/meetings/, then validate and apply it:

bash bin/ingest-graph.sh validate "<manifest>"
bash bin/ingest-graph.sh apply "<manifest>"

This manifest is the authoritative write contract for Meeting, Person, Decision, Finding, and provenance relationships. The route-specific Cypher later in this document is schema reference for legacy data and must not be executed as a second write path. Attendees resolve existing members by default; an unknown attendee requires an explicit source-scoped external identity and must never be auto-created as an organization member.

Step 0: Check Granola MCP connection

Granola data is accessed exclusively via MCP. The Granola MCP server must be configured in .claude/mcp.json and authenticated.

1. Check MCP availability: Use ToolSearch with query "granola" to check if Granola MCP tools are loaded.

2. If tools are available (e.g. list_meetings, get_meeting_transcript, get_meetings, query_granola_meetings): proceed to Step 0.5.

3. If tools are NOT available: check .claude/mcp.json for the Granola server config.

   • If config exists but tools aren't loaded: "Granola MCP is configured but not connected. Restart Claude Code to load the MCP server, then authenticate with /mcp → select granola → Authenticate."
   • If config doesn't exist: "Granola MCP not configured. Run /connect granola to set it up."
   • Stop in either case.

Step 0.5: Route subcommand

If $ARGUMENTS is backfill, jump to Backfill Mode (bottom of this document).

Step 1: Fetch unprocessed meetings

Read processed meeting IDs from state:

jq -r '.processed_meetings // {} | keys[]' .egregore-state.json

Call the list_meetings MCP tool to get recent meetings. Filter out already-processed IDs client-side.

If $ARGUMENTS is a search term (not "sync", not "backfill", and not empty), use the get_meetings MCP tool to search by title/content.

Step 2: Select meetings

Interactive mode (no arguments or search term): Present unprocessed meetings via AskUserQuestion:

question: "Which meeting should I process?"
header: "Meeting"
options:
  - label: "Meeting Title (Feb 12)"
    description: "With: Alice, Bob — 45 min"
  - label: "Another Meeting (Feb 11)"
    description: "With: Carol — 30 min"

Sync mode (/meeting sync): Process all unprocessed meetings. Show count:

Processing N unprocessed meetings...

Search mode (/meeting [search]): If exactly one match, use it. If multiple, present via AskUserQuestion. If none:

No meetings found matching "[search]".

Step 3: Fetch meeting data

Use the get_meeting_transcript MCP tool with the selected meeting's ID to retrieve the full transcript.

Use get_meetings or list_meetings for metadata (title, date, attendees) if not already available from Step 1.

The MCP tools return structured data directly — no JSON parsing or jq needed.

You now have: notes/summary, transcript, title, date, attendees.

Step 4: Load cross-meeting context (single batch call)

Run ALL 5 queries in a single bash bin/graph-batch.sh call (1 network round-trip instead of 5). Parse results[0] through results[4].

bash bin/graph-batch.sh '[
  {"statement": "MATCH (a:Artifact) WHERE a.origin STARTS WITH '"'"'granola:'"'"' AND a.created >= datetime() - duration('"'"'P30D'"'"') RETURN a.id, a.title, a.type, a.topics, a.meetingTitle, a.meetingDate, a.confidence ORDER BY a.created DESC LIMIT 15"},
  {"statement": "MATCH (a:Artifact) WHERE a.origin STARTS WITH '"'"'granola:'"'"' AND a.openQuestions IS NOT NULL RETURN a.title, a.openQuestions, a.meetingTitle, a.meetingDate ORDER BY a.created DESC LIMIT 10"},
  {"statement": "MATCH (a:Artifact) WHERE a.origin STARTS WITH '"'"'granola:'"'"' AND a.created >= datetime() - duration('"'"'P60D'"'"') UNWIND a.topics AS topic WITH topic, count(DISTINCT a.meetingTitle) AS meetingCount, collect(DISTINCT a.meetingTitle)[..5] AS meetings WHERE meetingCount >= 2 RETURN topic, meetingCount, meetings ORDER BY meetingCount DESC LIMIT 10"},
  {"statement": "MATCH (newer:Artifact)-[:SUPERSEDES]->(older:Artifact) WHERE newer.created >= datetime() - duration('"'"'P60D'"'"') RETURN newer.title AS current, older.title AS previous, newer.topics, newer.meetingTitle ORDER BY newer.created DESC LIMIT 10"},
  {"statement": "MATCH (q:Quest {status: '"'"'active'"'"'}) WHERE q.topics IS NOT NULL RETURN q.id, q.title, q.topics"}
]'

Result mapping:

• results[0] → Q1: Recent meeting artifacts (30d)
• results[1] → Q2: Open questions from previous meetings
• results[2] → Q3: Topic recurrence (60d)
• results[3] → Q4: Decision evolution chains
• results[4] → Q5: Active quests (for topic linking)

If the batch call fails, continue without cross-meeting context. It's enrichment, not required.

Step 5: Resolve attendees

Read the attendee map:

jq -r '.attendee_map // {}' .egregore-state.json

For each attendee from the meeting data, look up their graph name in the map. If an attendee is NOT in the map, auto-derive from email before asking:

1. If attendee has email: derive short name = email local part before @ (e.g., bartu@gizatech.xyz → bartu)
2. Check graph: MATCH (p:Person) WHERE p.name = $derived RETURN p — if person exists, auto-map silently
3. If name is clean (alphanumeric, not generic like "info" or "admin"): auto-map and show confirmation line (no AskUserQuestion): Auto-mapped: {full name} → {derived} (from email)
4. Only use AskUserQuestion if: no email available, email local part is ambiguous/generic, or multiple candidates exist

Save all new mappings to .egregore-state.json under attendee_map.

Use "me" for the user (from .egregore-state.json → name) and "them" for other attendees in speaker attribution.

Step 6: Intent harvesting

Per the Intent Harvesting section above. Use AskUserQuestion to understand what the user needs from this meeting before analysis begins.

Skip in sync mode — use adaptive defaults (balanced approach). Skip in backfill mode — historical, no user to ask. Skip if user already gave rich context with their invocation.

Step 7: Analyze

The free zone.

Apply the lenses that serve the user's intent and the material. Use the toolkit however you judge appropriate. Record what approach you chose.

Guidelines:

• If spawning sub-agents, use Task tool with model: "sonnet" and subagent_type: "general-purpose". Do NOT use run_in_background: true — run as standard parallel Task calls.
• Give each sub-agent: the relevant inputs for its lens, the output schema it needs to fill, and clear instructions.
• For inline analysis: read all inputs yourself and produce the output directly.
• For hybrid: scaffold inline, then spawn agents for specific lenses that need depth.
• Whatever approach you choose, the output must conform to the Enriched Artifact Schema and Meeting Intelligence Briefing format.

Step 8: Present proposal

Show the meta-analysis preview + merged extractions.

From "Weekly Sync — Feb 12" (Bob + Alice):
Approach: scaffold + inline analysis
Intent: full analysis
Tone: exploratory → decisive | Alignment: 0.85

  This meeting was really about settling the local transport question —
  everything else orbited that decision. The stdio choice unblocks MCP
  integration but creates a deferred problem for remote servers that
  nobody named yet.

  ◉ Decision: Use stdio transport for MCP servers           [0.9]
    "I tested both and stdio is way simpler" — them
    Conviction: assertion | Urgency: high
    Tradeoff: simpler local ↔ no remote support
    Open: how to handle remote MCP servers?
    Evolution: new position
    → mcp-integration

  ◉ Finding: Onboarding needs guided tour                   [0.9]
    "users drop off at step 4 without guidance" — me
    Conviction: data_backed | Urgency: medium
    → onboarding-flow

  ◉ Finding: Usage-based gating > tier gating                [0.5]
    * transcript-only — not in panel notes
    "gate by usage patterns not Claude tier" — them
    Conviction: exploration
    Open: what threshold?
    Evolution: shifted (was tier-based in Feb 8)
    → pricing-strategy

  Tensions:
    * "Usage gating" — stated as exploratory [0.5],
      but speaker energy suggests stronger conviction

  Actions:
    * Oz — write MCP auth spec by Friday
    * Bob — prototype guided tour by Monday

  Cross-refs: Decision (stdio) ← Finding (benchmark results)

Then present approval via AskUserQuestion (Enter confirms default):

question: "Proceed with these extractions?"
options:
  - label: "Yes, save"
    description: "Write files and index in graph"
  - label: "Edit"
    description: "Let me adjust before saving"
  - label: "Skip"
    description: "Skip this meeting"

Display rules:

• Approach: line showing what analysis method was used
• Intent: line showing what user asked for
• Tone summary + alignment score at the top
• Meta-analysis preview: 2-3 sentences covering "heart" + "what's new". Indented.
• [0.9] confidence score right-aligned after title
• Evidence quote in quotes with speaker attribution
• Conviction strength + urgency on one line
• * transcript-only flag for items not in panel (confidence < 0.7)
• Tradeoffs shown as X ↔ Y on one line when exactly 2, otherwise listed
• Open questions on separate Open: line
• Evolution context when available
• → quest-id for linked quests
• Tensions section after artifacts
• Actions listed separately at the end (not artifacts)
• Cross-refs shown if detected

Handle user response:

• "Yes, save" (default — triggers on Enter) → proceed to reflection checkpoint
• "Edit" → user modifies, then proceed
• "Skip" → skip this meeting, move to next

In sync mode: still present each meeting's proposal for confirmation.

Step 9: Reflection checkpoint

After the user approves, check if the analysis produced continuity signals worth reflecting on.

Check for reflection triggers:

Signal	Reflection prompt
Decision evolution with shifted or reversed trajectory	"Your position on {topic} has shifted from {old} to {new}. Is this intentional?"
Open thread from previous meeting addressed	"{Thread} from {meeting} appears to be resolved. Confirm?"
Open thread from previous meeting dropped	"{Thread} from {meeting} wasn't mentioned. Still relevant, or deprioritized?"
Meta-pattern detected	"Pattern detected: {description}. Is this becoming a principle?"
No meaningful signals	Skip reflection entirely

If a meaningful trigger exists, surface it via AskUserQuestion with specific options drawn from the actual data:

question: "{specific question from analysis}"
header: "Reflect"
options:
  - label: "{specific option 1}"
    description: "{what this means for the artifacts}"
  - label: "{specific option 2}"
    description: "{alternative interpretation}"
  - label: "Skip reflection"
    description: "Save artifacts as-is"

Handle response:

• Evolution confirmed → create additional decision artifact with SUPERSEDES relationship
• Compatible/clarification → add RELATES_TO relationship
• Open question → create finding with "Open: ..." prefix
• Skip → proceed without additional artifacts

If no continuity analysis was done or no meaningful signals, skip the reflection checkpoint entirely. Don't force it.