Meeting prep stack

You produce a one-page context pack for a marketing or GTM operator’s external meeting. The pack lands 30 minutes before the meeting starts. The operator reads it in three minutes and walks in informed.

Inputs to gather first

1. Meeting — calendar invite (title, time, attendees, description, link or location)
2. External attendees — name, email, company. The pipeline enriches each
3. Operator’s CRM access — for deal stage, related deals, last activity
4. Inbox history — last 90 days of threads with the attendees or company
5. Notes tool — Granola, Fireflies, Notion or Obsidian export with any prior meeting notes
6. LinkedIn source — official LinkedIn MCP, research-tool MCP, or manual paste
7. Operator’s relationship owner — who at their company owns the attendee or company relationship

If a source is missing, run the pipeline with what is available and explicitly mark the corresponding section as “not available”. Do not invent.

The pipeline (five phases)

Phase 1 — Pack spec

The pack has six sections:

1. Meeting metadata
2. Who you are meeting (one paragraph per external attendee)
3. What you have talked about (last 3 substantive interactions)
4. CRM context (deal stage, last activity, related deals)
5. Likely agenda (3-5 items)
6. The one thing (single most important context sentence)

Phase 2 — Attendee enrichment

For each external attendee, return JSON with name, current_role, tenure_at_current, background_summary (50-80 words), recent_signals (capped at 3, with source / date / summary), mutual_connections, prior_interactions_summary, confidence, uncertain_facts.

Rules:

• No invented biography. If LinkedIn does not say it, do not guess.
• recent_signals capped at 3.
• background_summary neutral tone, 50-80 words.
• Internal attendees get one line, name and role only. No enrichment.

Phase 3 — History aggregation

For each external attendee or company, return JSON with interactions (capped at 3, most recent first, with date / channel / summary / outstanding_followup), open_threads, last_substantive_decision, relationship_temperature (warm / neutral / cool / tense).

Rules:

• “Substantive” excludes scheduling, acknowledgments and cc threads.
• relationship_temperature reads from tone, not commercial state.

Phase 4 — CRM + agenda

Pull CRM record: company, primary_deal_stage, primary_deal_amount, primary_deal_close_date, last_activity_date, last_activity_summary, open_opportunities_count, related_deals (capped at 3), owner.

Draft agenda: 3-5 agenda items each with item (5-10 word noun phrase), rationale, estimated_minutes, decision_or_discussion. Plus the_one_thing (single sentence, sharp specific fact / decision / risk) and open_question_to_raise.

Rules:

• Sum of estimated_minutes equals meeting duration minus 10 minutes (buffer).
• the_one_thing is sharp. Not “be prepared”. A specific fact, decision, or risk.

Phase 5 — Pack assembly

Assemble into the template:

# {MEETING_TITLE}
{DATE}, {TIME_RANGE}, {LOCATION_OR_LINK}

## The one thing
{THE_ONE_THING}

## Who you are meeting
{EXTERNAL ATTENDEE PARAGRAPHS}

## What you have talked about
{LAST 3 INTERACTIONS}
{OPEN THREADS}

## CRM context
{CRM SUMMARY}

## Likely agenda
{NUMBERED ITEMS WITH ESTIMATED MINUTES}

## Open question to raise
{QUESTION OR 'none flagged'}

Output

Deliver as:

• Calendar event description, 30 minutes before start (default)
• Linked Google Doc or Notion page in the event
• Email to the operator at the same time
• Morning email at 06:30 if the operator has back-to-back meetings

Save to .lens/prep/{date}-{meeting-slug}.md. Downstream skills read it.

Evals

Before delivering:

• Length — pack reads in under 4 minutes. Cut padding.
• The-one-thing sharpness — single sentence, names a specific fact, decision, or risk. Not generic.
• Bio fidelity — every claim in background_summary traces to a source. Flag uncertain facts explicitly.
• History coverage — every substantive interaction in the last 90 days surfaces. Sample-check the source.
• Internal-attendee restraint — internal attendees are one line only. No full bios.

Failure modes to watch

• Invented biography — when LinkedIn is thin, the model fills in plausible-sounding facts. Hold the “no invented biography” rule. Use the uncertain_facts array to flag.
• Power dynamics missed — flag in open_question_to_raise whether the attendee can sign the deal or has to escalate.
• Old interactions feel current — every interaction is dated. Operator sees what is recent.
• Privacy by accident — the pack contains deal stages and sometimes compensation signals. Default delivery is the operator’s own event description, never a shared invite description.

Hand-off

• The pack feeds call-follow-up-loop for the post-meeting summary.
• daily-briefing-pipeline links to each pack from its “today’s meetings” section.
• personal-knowledge-base accumulates packs into a searchable interaction history.
• Open threads from agenda become tasks in inbox-to-task-pipeline if they need follow-up.

Save the prep pack to .lens/prep/{date}-{slug}.md.