Pre-Meeting Brief

v1.3 — Granola health-check simplified after the Granola public-API rewrite. Phase 0f no longer probes transcript content (the upcoming meeting has no transcript yet, and probing an arbitrary prior transcript adds latency without a clean health signal — it was producing false HOLLOW). HOLLOW classification removed for Granola entirely; Granola is now OK / DEGRADED / UNAVAILABLE only. Phase 1c fallback clarified to hunt PRIOR project meetings for context, not the upcoming one. Source-confidence table, circuit breaker, and error-handling table aligned. v1.2 — Generates a structured pre-meeting briefing from project communications. Run from a project directory.

Pulls context from WhatsApp, Gmail, Granola, Google Docs, and local files. Surfaces decisions required, open loops, and preparation needs. Optionally emails the brief to attendees.

Prerequisites

Required:

• Gmail MCP — for searching and reading message context
• Google Calendar MCP — for identifying the target meeting and attendees

Recommended:

• Granola MCP — for pulling meeting transcripts
• WhatsApp MCP — for picking up project group threads
• A project .claude/CLAUDE.md with team roster, group JIDs, dashboard doc ID, and folder paths

Optional:

• A send-email helper script if you want one-step send (otherwise the skill creates a Gmail draft)

First-Time Setup

1. Create a project config at [project-root]/.claude/CLAUDE.md with at minimum:

   # Project Name
   
   ## Team
   - [Name] <email@example.com> — PI
   - [Name] <email@example.com> — Research Manager
   - [Name] <email@example.com> — RA
   
   ## Folders
   - transcripts: `transcripts/`
   - weekly_reviews: `weekly-reviews/`
   
   ## Config
   ```yaml
   project_keywords: [keyword1, keyword2]
   exclude_keywords: [unrelated_term]
   whatsapp_groups: ["Project Group Name"]
   google_doc_id: "YOUR_DOC_ID"
   granola_folder: "Project Folder Name"   # optional

2. Create a transcripts folder in your project directory:

   mkdir -p transcripts/

3. (Optional) Configure a routing file at ~/.claude-assistant/config/granola-routing.json if you have multiple projects pulling Granola transcripts. The skill uses this to list known projects when it can't find a config.

Arguments

$ARGUMENTS can include (combine freely):

• (none) — next upcoming meeting with project context
• [keyword] — filter by meeting title
• nosend — skip the email phase
• email — include email draft (prompt to send)
• light — force Light depth (overrides auto-detection)
• full — force Full depth (overrides auto-detection)
• since:YYYY-MM-DD — override lookback start date

CRITICAL: No Permission Prompts

Do NOT use Task agents or ToolSearch. All required tools are pre-approved. Call them directly:

• Read, Write, Bash, Glob, Grep
• mcp__google_workspace__get_events, mcp__google_workspace__search_gmail_messages, mcp__google_workspace__get_gmail_message_content
• mcp__google_workspace__get_doc_content, mcp__google_workspace__draft_gmail_message
• mcp__granola-api__search_granola_events, mcp__granola-api__get_granola_document, mcp__granola-api__get_granola_transcript
• mcp__whatsapp__list_messages, mcp__whatsapp__search_messages, mcp__whatsapp__list_chats

Risk escalation:

• Data gathering (all reads/searches): Auto
• Briefing display: Auto
• Email draft display: Batch-confirm (show draft + recipients)
• Email send: Per-item confirm (user must say "send")

Instructions

Phase 0: Identify Meeting + Pre-flight

0a. Read project config: Read $(pwd)/.claude/CLAUDE.md. Extract:

• whatsapp_groups — group names (look up JIDs via sqlite3 ~/whatsapp-mcp-ts/data/whatsapp.db "SELECT jid, name FROM chats WHERE name LIKE '%[group_name]%'")
• google_doc_id — project dashboard doc
• team roster — all role groups (pis, research_managers, field_team, ras, etc.)
• project_keywords, exclude_keywords
• folders.transcripts, folders.weekly_reviews
• granola_folder or granola_folders — Granola folder name(s) for transcript fetching
• pre_meeting_brief section (if present — optional config overrides)

Extract fields from inside markdown code fences if present.

If no .claude/CLAUDE.md found: Stop. Read ~/.claude-assistant/config/granola-routing.json (if present) and list known project directories:

No .claude/CLAUDE.md found in this directory.
Known project directories:
[list from granola-routing.json routes]
cd into a project directory and re-run /pre-meeting-brief.

0b. Identify target meeting:

1. Fetch today + tomorrow events via mcp__google_workspace__get_events (all calendars)
2. If [keyword] argument: filter by title contains keyword (case-insensitive)
3. Else: match events where attendee emails overlap project team roster, OR title contains any project_keyword
4. If 0 matches: "No matching meetings found today or tomorrow. Specify a keyword or check your calendar."
5. If 2+ matches: show numbered list, ask user to pick
6. If 1 match: proceed

0c. Determine depth + funder flag:

• Auto-Light if: meeting ≤30 min AND calendar title contains "standup", "check-in", "weekly sync", "1:1", "one-on-one" AND no full argument
• Auto-Full otherwise
• Override with light or full argument
• Funder flag: set if any attendee matches known funder contacts from a configured donor profiles folder (e.g., ~/Documents/donors/) or if the project config lists them as funders. If set, donor profile lookup is included in Phase 1.

0d. Determine lookback window: Priority order:

1. since:YYYY-MM-DD argument (explicit override)
2. Most recent weekly-review-*.md in folders.weekly_reviews
3. Fallback: 14 days ago

Display in output header: "Lookback since [date] ([anchor: weekly review / 14-day default])"

0e. Populate local transcripts: If granola_folder or granola_folders is configured in .claude/CLAUDE.md and you have a transcript fetcher (e.g., granola-fetch): Run the fetcher once per configured folder for the lookback window. This ensures local transcript files are up-to-date before Phase 1. If neither field exists: skip. If the fetcher fails: note the error, continue (Phase 1 will use existing local files + MCP fallback).

0f. Pre-flight health checks:

Run all checks silently. Results feed into the Phase 2 banner. No user interaction.

Fire in parallel:

1. WhatsApp SQLite: sqlite3 ~/whatsapp-mcp-ts/data/whatsapp.db "SELECT max(timestamp), count(*) FROM messages WHERE chat_jid IN ([configured JIDs]) AND timestamp > [lookback_epoch]"
2. Gmail: one test search with project keywords — check auth + result count
3. Granola: search_granola_events with project keywords to check that the API + key are working AND that the project has prior meetings indexed. Do NOT probe transcript content here — the upcoming meeting has no transcript yet, and probing an arbitrary prior-meeting transcript adds latency without a clean health signal.
   • Events found → OK
   • No events found → DEGRADED ("no prior project meetings indexed in Granola")
   • Auth error / unreachable / 401 → UNAVAILABLE (re-auth Granola per your provider's instructions)
4. Google Doc: Read first 500 chars via mcp__google_workspace__get_doc_content (if google_doc_id configured). If readable but body < 100 chars → HOLLOW.
5. Local files: glob transcripts folder — check existence

Source confidence thresholds:

HOLLOW counts as DEGRADED for the minimum-signal gate.

Minimum-signal gate: Require ≥2 sources at OK or DEGRADED/HOLLOW. If only 1 source is functional → show a raw data dump with: "Insufficient data for structured brief. Only [source] available. Raw findings below." Do not present as a structured briefing.

Phase 1: Gather Data (inline parallel MCP calls)

Per-source call limits (Full depth):

When a source hits its cap: stop immediately. Note "[SOURCE] capped" in the source banner. Proceed to synthesis with what was gathered. Do NOT debug or retry.

Granola Phase 1 check: before making any Phase 1 Granola calls, check if Phase 0f flagged Granola as DEGRADED or UNAVAILABLE. If UNAVAILABLE, skip Granola in Phase 1 entirely. If DEGRADED (no prior project meetings), skip the Phase 1c MCP fallback but still attempt local-file reads.

Full depth (target 50, warn at 80 tool calls):

Fire these in parallel where possible:

• 1a. WhatsApp (if OK/DEGRADED):
  • MCP path (if OK): Query via mcp__whatsapp__list_messages. Filter received messages in-context: discard any message where timestamp (ISO 8601 string, format: YYYY-MM-DDTHH:mm:ss.sssZ) is earlier than lookback_start. The column is already ISO 8601 text — compare as strings: timestamp >= lookback_start_iso.
  • SQLite path (if DEGRADED): Query with WHERE timestamp >= '[lookback_start_iso]' (ISO string comparison works correctly on the messages table schema).
  • Post-process (both paths): Group messages by calendar day. Insert date headers between day groups when passing to synthesis:

    --- March 18, 2026 ---
    [messages from that day]
    --- March 17, 2026 ---
    [messages from that day]

    Date headers appear in synthesizer context only — not in the output sections. Do not include messages earlier than lookback_start in the context passed to synthesis.

• 1b. Gmail — 4-search priority stack, fire in parallel:

  • Search 1 (PI priority, 2d): from:({pi_emails}) newer_than:2d — read ALL results. These are highest-priority input regardless of recency within the lookback window.
  • Search 2 (attendee recent, 3d): (from:{attendee_emails}) newer_than:3d
  • Search 3 (team + keywords, full lookback): (from:{team_emails}) ({keywords}) newer_than:{lookback_days}d
  • Search 4 (WhatsApp compensation — only if WhatsApp DEGRADED/UNAVAILABLE): (from:{team_emails}) newer_than:{2*lookback_days}d
  • Read content for top 10-15 relevant messages. Prioritize Search 1-2 results.
  • Gmail cap: 8 total calls (4 searches + up to 4 content reads) max. Agenda detection (Phase 1g) reuses gathered content — does not fire new searches.

  Source ranking in synthesis: When multiple sources cover the same topic, weight: recent PI email (Search 1-2) > posted agenda > recent WhatsApp > older transcripts/email. When sources conflict: note the conflict rather than silently preferring one.

• 1c. Granola transcripts: Primary: Read local .txt files from the transcripts folder (populated by Phase 0e). Apply hollow-transcript check: FULL (≥5,000 chars + conversational speaker content) vs HOLLOW (<5,000 chars, starts with ###, or bullet-only without speaker attribution). Exclude hollow transcripts from substantive use. Fallback (only if no local files AND prior project meetings exist — for upcoming meetings the current meeting has no transcript yet, so this fallback fetches PRIOR project meetings for context, not the upcoming one): Call list_granola_documents to get Granola UUIDs, matching to prior project meetings by title/date. Use those UUIDs with get_granola_transcript — never use Calendar event IDs from search_granola_events directly. (Note: Granola transcript reads are the most likely source of budget overruns — cap at 3 transcripts.)
• 1d. Google Doc dashboard (if configured + OK): Read project dashboard tab for status, metrics, action items via mcp__google_workspace__get_doc_content.
• 1e. Local files: Glob transcripts folder for new files since lookback. Read HANDOFF.md, TODO.md, PROJECT_STATE*.md if present (silent skip if not found).
• 1f. Donor profile (if funder flag set): Read the relevant donor profile from your configured donor folder (e.g., ~/Documents/donors/[slug].md).
• 1g. Agenda detection (runs on already-gathered data — no new tool calls): Scan retrieved WhatsApp messages and Gmail content for agenda signals:
  • Keywords: "agenda", "topics to cover", "topics for today", "order of business"
  • Structure signals: numbered or bulleted list of meeting topics
  • Look within 48h of meeting time and from meeting participants
  • If found: extract ordered list of agenda items. Note poster and timestamp. Store as detected_agenda: [items].
  • If not found: detected_agenda: null.

Light depth (8-15 tool calls, 15-30 sec):

• Last Granola transcript only (if exists)
• Gmail: 1 search (team roster emails, last 14 days), read top 5 messages
• WhatsApp: messages in project group since last meeting
• Skip Google Doc, local files, donor profile

Phase 2: Synthesize

2a. Build one-line source banner (degraded sources only):

• If all OK: Sources: all healthy
• If any degraded/unavailable: ⚠ Sources: WhatsApp DEGRADED (MCP filter issue, using SQLite) | Granola UNAVAILABLE (auth expired)
• Suppress healthy sources from the banner.

2b. Cross-reference action items: If HANDOFF.md, weekly review, or previous Granola transcript found, extract action items. For each, search ALL gathered data for evidence of completion:

• DONE: Explicitly marked complete in 2+ independent sources
• LIKELY DONE: Appears resolved in 1 source, no contradicting evidence
• NO EVIDENCE: Appears open in any source, or only in 1 source as open
• CONTESTED: Different sources give conflicting status

Surface only NO EVIDENCE and CONTESTED items in the brief. Suppress DONE and LIKELY DONE.

2c. Organize and classify:

• If detected_agenda is not null: Use BY AGENDA ITEM format. For each agenda item, find relevant gathered context and tag each claim as DECIDE/UPDATE/OPEN LOOP. Note: the agenda item itself has no confidence tag — the claims within it get confidence tags. Append "Items not on agenda" section for any DECIDE/OPEN LOOP items not covered by the agenda.
• If detected_agenda is null: Use flat classification:
  • DECIDE: Unresolved thread with "what do you think?" / "can we proceed?" / "need your sign-off"; overdue action item assigned to you; conflicting proposals; budget/timeline questions without answers
  • UPDATE: Status reports, completed action items, FYI threads, meeting logistics, confirmations
  • OPEN LOOP: Items others owe you that are overdue or unacknowledged

2d. Per-attendee framing (Full only): For each attendee: compile "What I owe them" and "What they owe me" from cross-referenced action items.

2e. Identify preparation needs (Full only): From gathered data: documents to share, slides to review, draft decisions you need to have ready.

2f. Claim confidence tags (Full depth only): For each item in DECIDE and OPEN LOOPS, assess and tag:

• VERIFIED: Corroborated by 2+ independent channels (independent = different communication channels, e.g., email + WhatsApp; NOT two retrieval paths for the same event, e.g., Granola + WhatsApp transcript of the same meeting).
• SINGLE-SOURCE [channel, date]: Explicitly stated in exactly one channel.
• INFERRED: Derived from absence, pattern, or combination — not explicitly stated in any source.

Rules:

• DECIDE items tagged INFERRED only → suppress the item. Write "No decisions identified" if DECIDE section would otherwise be empty.
• OPEN LOOP items tagged INFERRED → include but mark visibly.
• Cross-check any claim about "[you] said/did X" against your sent emails before including it. If not found in sent email, downgrade to SINGLE-SOURCE or INFERRED.
• Temporal attribution: when attributing a claim, use the timestamp of the original event described, not the message timestamp. If a message references a past decision without a dated original event, tag INFERRED with note: "date of original event unknown."

Phase 3: Present Briefing

Full depth: Full terminal template with all sections (DECISIONS REQUIRED, OPEN LOOPS, PER ATTENDEE, PREPARE, KEY UPDATES, SUGGESTED FOCUS, KNOWN BLIND SPOTS).

Light depth: 3-4 bullet format (last met, I owe them, they owe me, open question).

After display (Full only):

• If ≥1 DECIDE item or ≥1 "I owe them" item: Send to attendees? (Send / Edit / Save / Skip)
• Otherwise: Save locally? (Save / Skip) — don't offer email for meetings with no actionable content.

If nosend argument: skip email prompt entirely. If email argument: always prompt for email regardless of content.

Phase 4: Distribute (if email requested)

4a. Draft email: The email is a document delivery to attendees, not a summary for yourself.

KEEP (include in full, reframe as noted):

• DECISIONS / BY AGENDA ITEM content → reframe as "Topics to cover" or by agenda item
• KEY UPDATES → keep as "Recent updates"
• SUGGESTED FOCUS → keep as "Suggested focus"
• Any materials circulated since last meeting → list under "Materials shared"
• Do NOT include INFERRED items in the email even if tagged — external recipients should only see VERIFIED or SINGLE-SOURCE claims.

REMOVE (internal to you only):

• Source banner (⚠ Sources: ...)
• OPEN LOOPS section (items others owe you — not the attendees' problem)
• PREPARE section
• PER ATTENDEE framing
• KNOWN BLIND SPOTS
• Any claim tagged INFERRED

Length: The email should contain all the same substantive content as the terminal brief minus the removed sections. Do not compress or summarize — present the full content. A 40-line terminal brief produces a ~25-line email, not a 10-line email.

Use a direct, professional tone. Avoid AI-sounding language ("delve," "leverage," "multifaceted," etc.). Recipients come from the calendar invite attendee list.

4b. Show draft + recipients for confirmation:

────────────────────
PRE-MEETING EMAIL
────────────────────

To: [attendee names + emails from calendar invite]
Subject: Ahead of [meeting title] — [date]

[draft]

→ Send / Edit / Skip

4c. Send:

Option A: Gmail MCP draft (simple) Create a draft via mcp__google_workspace__draft_gmail_message. Report: "Draft created in Gmail — open Gmail to review and send."

Option B: Custom send script (advanced) If you have a send-email script at ~/.claude-assistant/scripts/send-email.py:

cat > /tmp/pre-meeting-body.txt << 'BODY'
[email body]
BODY

python3 ~/.claude-assistant/scripts/send-email.py \
  --to "recipient1@email.com" --to "recipient2@email.com" \
  --subject "Ahead of [meeting title] — [date]" \
  --body-file /tmp/pre-meeting-body.txt

rm -f /tmp/pre-meeting-body.txt

4d. Fallback: If send fails, fall back to mcp__google_workspace__draft_gmail_message. Note "Saved as Gmail draft."

4e. Save locally (if requested or save argument): Write to [transcripts_folder]/pre-meeting-brief-YYYY-MM-DD.md.

Phase 5: Log Performance

echo "$(date +%Y-%m-%d),pre-meeting-brief,TOOL_CALLS,NOTES" >> ~/.claude-assistant/logs/skill-performance.csv

Replace TOOL_CALLS with the exact count (each MCP call, Read, Write, Glob, Grep, Bash = 1). Replace NOTES with a summary (e.g., team-sync-full-3-decide).

After logging: "Run /post-meeting after this call to capture outcomes."

Customization Points

• Transcript source: Swap Granola for Otter, Fireflies, or any tool that exports .txt transcripts to your transcripts folder.
• WhatsApp DB path: The default ~/whatsapp-mcp-ts/data/whatsapp.db reflects one MCP install. Update the SQLite paths if your bridge stores data elsewhere.
• Donor profile folder: The funder-flag lookup expects a [slug].md file per funder in a configured folder (e.g., ~/Documents/donors/). Skip the funder branch entirely if you don't track funders this way.
• Email sending: Default is Gmail draft (Option A). Replace with a send script (Option B) if you have one configured.
• Routing config: ~/.claude-assistant/config/granola-routing.json is optional — used only to suggest known project directories when the skill is run outside one.

Error Handling

Performance Budget

Advisory threshold: 80 tool calls. Per-source caps bound tool call distribution: Gmail 8, Granola 5, WhatsApp 5, Doc 2, Local 3, plus Phase 0 overhead (~8) = ~31 calls minimum floor. Target 50 total.

Examples

/pre-meeting-brief                    # Next meeting matching project context
/pre-meeting-brief Carnegie           # Meeting with "Carnegie" in title
/pre-meeting-brief light              # Force light depth
/pre-meeting-brief full email         # Force full + prompt email
/pre-meeting-brief nosend             # Show brief only, no email
/pre-meeting-brief since:2026-03-01   # Override lookback start