How I AIHow this PM uses MCPs to automate his meeting prep, CRM updates, and customer feedback synthesis
CHAPTERS
- 0:00 – 2:41
MCPs demystified: app connectors for your AI tools
Claire and Reid open by reframing MCPs in plain language: they’re essentially app integrations/connectors that let AI tools both read knowledge from your apps and take actions inside them. They outline the two core motivations—accessing context and executing work—setting up the practical demos that follow.
- •Don’t get stuck on the term “MCP”—think “AI app connectors”
- •Two primary use cases: read knowledge in apps + take actions in apps
- •MCPs work across popular AI clients (Claude, ChatGPT, Cursor, etc.)
- •Why MCPs feel hyped but underused: conceptual friction and setup complexity
- 2:41 – 4:05
Zapier’s MCP approach: 8,000 apps and tool collections per client
Reid explains how Zapier exposes a massive library of searches/actions via MCP and why Zapier’s server approach is more like a platform for building multiple servers. The key idea is creating curated toolsets tailored to a specific AI client or agent so the model has the right capabilities without excess access.
- •Zapier MCP exposes ~8,000 apps and tens of thousands of actions/searches
- •Create multiple MCP “servers” (tool bundles) for different contexts
- •One URL connector can grant an AI client access to a curated toolset
- •Tool restrictions matter: limit access to specific docs, databases, notebooks
- 4:05 – 9:00
Configuring tools responsibly: scoping access and reducing tool confusion
They discuss practical configuration details: restricting which databases, notebooks, or records an AI can touch and why that improves safety and reliability. Claire highlights a common pain point—tool selection conflicts—arguing that better prioritization and control will be increasingly important as MCP adoption grows.
- •Fine-grained scoping: restrict Coda tables, Evernote notebooks, etc.
- •Custom tool bundles reduce noise and mitigate accidental actions
- •Tool-calling ambiguity is real: models can pick the wrong tool
- •Emerging need: tool priority/selection controls in MCP clients
- 9:00 – 12:05
Claude Projects as a tool-usage “playbook” for MCP workflows
Reid shares a key tactic: using Claude Projects not just for knowledge, but for detailed operational instructions on how tools should be used, in what order, and how data should be written back. This turns a general-purpose model into a more consistent operator for repeated workflows like CRM updates.
- •Claude Projects can store step-by-step tool calling instructions
- •Define sequencing: which tools to use first, second, third
- •Specify mapping: what data goes into which fields/records
- •Switch projects by task type (e.g., “CRM project”) to improve consistency
- 12:05 – 15:25
Daily meeting prep automation: calendar scan + CRM/company intelligence
Reid describes a daily planning workflow that reviews his calendar and automatically enriches each meeting with context—who the person is, company usage, prior interactions, and other internal data. The goal is to stop showing up to meetings without context and to help customer-facing teams be better prepared by default.
- •Automated “daily plan” built from calendar events
- •Lookups: person identity, company/app usage, prior touchpoints
- •Output: a single daily brief with embedded research
- •Reusable pattern for sales teams and customer-facing roles
- 15:25 – 18:15
Post-meeting notes management: from Granola notes to structured records
They walk through Reid’s pain point: post-meeting note logging is tedious and easy to skip. Using MCP-enabled tools plus project instructions, Claude can check for existing records, search internal sources (like Glean/Slack knowledge), and create or update structured entries in systems like Coda and HubSpot.
- •Meeting notes often don’t get logged due to friction
- •Workflow: check if interview/record exists → enrich → create/update
- •Use internal search (e.g., Glean) to pull relevant context
- •Write outcomes back to Coda/CRM: attendees, next steps, opportunity details
- 18:15 – 20:04
Agentic instructions vs deterministic Zaps: brittleness, speed, and “where you work”
Claire contrasts her deterministic Zapier workflow builder approach with Reid’s agentic, natural-language instruction approach in Claude. Reid explains trade-offs: agentic flows can be more flexible but may hit time limits, while deterministic workflows handle long-running steps better; the big win for MCPs is bringing actions into the AI tools people already use.
- •Deterministic workflows: explicit branching and reliability for long tasks
- •Agentic instructions: flexible but constrained by latency/timeouts
- •Core value of MCP: meet users inside their preferred AI client/IDE
- •Enterprise angle: role-based toolsets provisioned by admins
- 20:04 – 23:10
Idea jammer: using MCP tool access to challenge and refine concepts
Reid briefly shares an “idea jammer” Claude Project wired to his tables and references. The project helps him explore whether ideas were tried before, where they might fit, and applies prompting methods to challenge assumptions—showing MCP value beyond pure admin work.
- •Dedicated project for brainstorming and structured ideation
- •Researches prior similar ideas and relevant internal context
- •Uses prompting frameworks to challenge thinking and assumptions
- •Example of MCP helping creative/strategic work, not just ops tasks
- 23:10 – 25:05
Customer interview prep workflow: Databricks lookup + Coda briefing page
For interview preparation, Reid relies on a more deterministic workflow because the underlying research lookup can take time. The flow pulls enriched customer/company context (e.g., from Databricks), summarizes it, and appends it to a Coda page so he can enter interviews fully informed—even when the lead came from loosely tracked sources like LinkedIn.
- •Problem: interviews booked from many sources → little context at call time
- •Automated lookup by email to generate a rich customer/company brief
- •Write/append the brief into a Coda page used for prep and notes
- •Deterministic workflow chosen because the data retrieval can be slow
- 25:05 – 29:16
Why Gemini is in the loop: file-native processing for PDFs/HTML
They unpack a model-choice nuance: Gemini performs especially well on file-based inputs. Reid’s data output comes as HTML/PDF-like artifacts, so he converts HTML to a file to reduce tokens and improve processing quality, then uses Gemini in the workflow to interpret and summarize the document reliably.
- •Gemini is often strong for file-based/large-context inputs
- •Reid’s source output is HTML (converted to a file for efficiency)
- •File conversion can reduce token usage and improve throughput
- •Practical takeaway: pick models based on data format, not just preference
- 29:16 – 31:48
Creating a virtuous cycle: turning support tickets and transcripts into FAQs
Reid describes a system that keeps customer-facing knowledge fresh without constant manual upkeep. Each closed support ticket or chatbot transcript is analyzed to extract the core question and solution; if it’s missing from the knowledge base, the system proposes a new FAQ entry for human review and then publishes it to the bot’s source database.
- •Knowledge bases rot without continuous updates
- •Auto-analyze closed tickets/transcripts to extract FAQ candidates
- •Check if answer already exists; propose net-new entries when missing
- •Human-in-the-loop approval step before publishing to bot knowledge
- 31:48 – 33:03
“Run ChatGPT in your sleep”: a framework for brainstorming high-leverage automation
Claire and Reid share mental models for finding strong AI use cases: imagine an infinite, perfect team or imagine running AI while you sleep. The point is to move beyond “do the same work faster” toward “raise quality and coverage,” like systematically reviewing every support interaction to improve documentation.
- •Use-case prompt: “If you could run ChatGPT in your sleep, what would it do?”
- •Alternative prompt: “What would an infinite perfect team do?”
- •Shift from speed to quality: broader coverage, fewer missed improvements
- •Product idea: Mad Libs-style prompts to help users discover workflows
- 33:03 – 37:16
Personal workflows: family calendar from a photo + kid music with Suno
In the lightning round, Reid shares personal automations that deliver real-life value. He uses a Claude Project plus Zapier MCP to translate photos of a physical family calendar into correctly blocked Google Calendar events (including travel buffers), and he uses Claude + Suno to generate custom kids’ songs—both for fun and to teach kids prompt specificity.
- •Family calendar workflow: photo → interpret → create/update Google Calendar events
- •Adds practical constraints: commute buffers during work hours
- •Suno + Claude: personalized kids’ songs with learn-by-doing prompting
- •Creative reuse: turning transcripts/decks into memorable songs
- 37:16 – 40:28
NotebookLM interview prep: personalized audio briefings for a job search
Reid closes with a high-impact personal use case: preparing his wife for interviews using NotebookLM. He gathers company/job materials, generates an “audio overview” tailored to her background, and she listens before interviews—improving her readiness and perceived domain knowledge.
- •NotebookLM used for learning and interview preparation
- •Inputs: careers page, job description, additional company context
- •Output: personalized audio briefing that addresses the candidate directly
- •Practical benefit: stronger interview performance via tailored context
