The exact AI playbook (MCPs, GPTs, Granola) that saved ElevenLabs $100k+ & helps them ship daily

Youtube URL: https://youtu.be/5Byg-9K8JnM

Host(s): Claire Vo (How I AI)

Guest(s): Luke Harris, Head of Growth at ElevenLabs

Podcast Overview and Key Segments

Overall Summary

Claire Vo hosts Luke Harris to unpack ElevenLabs’ practical AI playbook for shipping daily and cutting costs. Luke shows a live case study workflow using Granola and a custom GPT to turn short interviews into polished posts and tweets. He explains how he replaced a $40k/year localization SaaS and $100k+ in agency spend with a simple LLM-powered GitHub Action and language-specific prompts. He also demos an open-source WhatsApp MCP that lets Claude summarize, search, and send messages from WhatsApp, then chain tools (like ElevenLabs voice) for voice roundups or on-the-fly agents. Core lessons: make everything a launch, edit the underlying prompt not the output, stitch tools via orchestration, and rethink build vs buy when AI quality fast outpaces legacy SaaS.

Reference

• MCP (Model Context Protocol): A standard that exposes local or remote tools to AI agents so they can act (query data, send messages, generate media).
• GPTs: Custom versions of ChatGPT with instructions, files, and tools, reusable by teams.
• Granola: AI meeting notes/transcription tool that outputs summaries and full transcripts.
• Cursor: AI-native code editor for fast prototyping and refactors with model help.
• GitHub Action: Workflow that runs on repo events (e.g., on push) to automate tasks.
• Payload CMS: Headless CMS used to manage website content.
• Agentic workflows: Systems where AI plans, calls tools, and executes steps with minimal human glue.
• Human-in-the-loop: People review or approve AI output at key checkpoints.
• WhatsApp MCP / Whatsmeow: Unofficial bridge that mirrors WhatsApp Web to fetch/send messages locally using the Whatsmeow library.
• SQLite: Lightweight local database to store messages.
• Orchestration (Conductor by Orkes): A platform to coordinate multi-step workflows, humans, systems, and agents.

Key Topics

AI CMO and “Everything is a Launch”

Luke predicts a near-term shift where the AI CMO automates full launch pipelines: value props, messaging, assets, blog posts, tweets, ads, and landing page tests. Each feature ships with end-to-end orchestration and evergreen channels. The goal is speed and consistency. AI handles copy, assets, and distribution setup. Humans steer strategy and quality. ElevenLabs treats every feature and case study like a launch. They aim to convert product changes into campaigns, iterate fast, and keep momentum. The lesson: your product only matters if people use it. Market velocity beats tool velocity.

Case Study Engine: Granola + Custom GPT

Workflow: record a 3–5 minute Zoom-style interview. Granola creates a sharp summary and a full transcript. Paste both into a company GPT that is primed with tone, examples, and strict formatting rules. Output: a clean case study plus a tweet thread with asset placeholders. Iterate by refining the prompt, not the output. Add internal links or product facts as needed. Automate scheduling with a CRM trigger (close-won → Calendly). This turns sporadic case studies into a steady stream. The result is consistent storytelling, faster content cycles, and quotes pulled from raw transcripts.

Killing a $40k SaaS and $100k Agencies with LLMs

Problem: localization tools were pricey, rigid, and slow. Agencies varied in quality and took days. Solution: a small server + GitHub Action that reads translation keys, calls an LLM with a language-specific prompt (tone, glossary, brand rules), and writes translations back to GitHub or the CMS. Reviews are “human-in-the-loop” only for sensitive pages. Time-to-translation goes from days to instant. Cost goes to near-zero. Quality rises due to prompt control and consistent tone. Key lesson: if a vendor limits prompt control, quality stalls. Build when AI is “good enough” and improving.

WhatsApp MCP: Your Messages, Summarized and Actionable

Luke open-sourced a WhatsApp MCP that works by mirroring WhatsApp Web. It downloads messages once to a local SQLite DB and exposes tools to Claude. Use it to summarize group chats, track trends, search topics, and even send messages or voice notes. You can chain it with other MCPs, like ElevenLabs, to create daily voice roundups. Data stays local. It updates when running. It is unofficial, so there’s a ban risk, but it minimizes calls by syncing once. This turns your WhatsApp into a searchable knowledge feed for content ideas, market intel, and personal ops.

Prompt Engineering: Edit the Prompt, Not the Output

Luke stresses refining the underlying prompt when you notice repeat issues. Want stronger headers, more numbers, or quotes? Bake it into the instructions. Provide tone guides, good examples, and even bad examples for contrast. Use American English if that is your standard. Specify formatting for tweets vs blog posts vs ads. This avoids manual edits later. It also scales quality across content types and teams.

Voice Modality: New UX and Scalable Ops

Voice unlocks new experiences and operational scale. Examples: interactive tutors in education (chess.com’s Professor Wolf), product keynote prototypes using ElevenLabs Studio voices, and multilingual support agents. Voice adds engagement and clarity, especially for complex topics. On the ops side, voice agents can extend support to new languages, run research calls, or collect data at scale. Combine with translation prompts to match local styles and norms. Human review remains useful for sensitive cases.

Key Themes

Automation as a Growth Edge

Automate launches end-to-end. Codify messaging, tone, and assets. Use CRM triggers to schedule customer interviews. Convert transcripts into multi-channel content. Test evergreen channels like Google Ads fast. The compounding effect is throughput. You publish more, learn faster, and stay visible. Quote:

• “Everything is a launch.”
• “The thing I’m really excited [about] for the AI CMO is translating every feature into your entire launch process.”

Build vs Buy in the AI Era

AI cut build time and raised quality. For localization, the “buy” path locked prompts and forced agencies. The “build” path added control, speed, and savings. If a tool blocks prompt editing, expect worse quality over time. Bet on AI getting cheaper and better. Quote:

• “This saved us $40,000 a year for the tool… Over $100,000 in agency costs.”
• “I did the first 90% in one day in Cursor.”

Agentic Workflows and MCPs

Static automations break when tasks change. Agentic workflows let models select tools and improvise. MCPs expose local tools like WhatsApp, voice, or CMS. Agents can summarize chats, draft tweets, or call users and run interviews. This reduces glue work and unlocks new flows. Quote:

• “The really cool thing about these chat-based MCP tools is it can be much more… able to deal with higher-level tasks.”
• “You can on the spot come up with these agents… which can then do these tasks for you.”

Prompt Governance and Brand Consistency

Good outcomes depend on good prompts. Centralize tone of voice, examples, glossary, and language nuances. Keep per-language prompt files. Enforce formatting rules by content type. Edit prompts over time. This builds brand consistency at scale. Quote:

• “When you’re editing… try and edit the underlying prompt rather than the actual output.”
• “Give it as much context as possible.”

Local-First, Privacy, and Speed

The WhatsApp MCP runs locally. Messages are stored in SQLite. You get speed, privacy, and reduced ban risk. Local-first is a viable pattern for sensitive data. It also allows multi-tool chaining without shipping data across clouds. Quote:

• “It downloads all your messages onto your local computer… Then you can keep querying it.”
• “None of this stuff is going to the cloud.”

Key Actionable Advise

Key Problem

Inconsistent launch execution and slow content output.

• Solution Build a “Case Study Engine” with Granola + a custom GPT.
• How to Implement Set a CRM trigger (close-won) to auto-send a Calendly link. Record 3–5 minute interviews. Feed Granola’s summary and transcript to a GPT primed with tone, examples, and formatting rules. Generate blog posts and tweet threads. Refine the prompt, not outputs.
• Risks to be aware of Model hallucinations. Incorrect product facts. Add human review for sensitive claims and link to source docs.

Key Problem

High localization costs and low quality.

• Solution Replace vendors with a GitHub Action + LLM translation per language.
• How to Implement Store copy in code (JSON) and CMS. On push, run an action that sends keys to an LLM with per-language prompts (tone, glossary, examples). Write translations back. Review only critical pages.
• Risks to be aware of Legal or pricing pages need human checks. Brand drift if prompts are weak. Track versioning and rollback.

Key Problem

Information overload in WhatsApp groups.

• Solution Use a local WhatsApp MCP to summarize, search, and act.
• How to Implement Run the open-source MCP. Scan the WhatsApp Web QR. Let Claude index and summarize threads, extract trends, and draft posts. Chain with voice for daily roundups.
• Risks to be aware of Unofficial bridge risk. Keep syncs local. Respect privacy and group norms.

Key Problem

Manual edits and tone inconsistency across content.

• Solution Create a strict prompt governance layer.
• How to Implement Maintain a central tone guide, examples (good and bad), and format rules per content type. Enforce American English or locale rules. Update prompts as feedback arises.
• Risks to be aware of Overfitting tone. Stale examples. Review and refresh quarterly.

Key Problem

Limited support capacity and lack of multilingual coverage.

• Solution Spin up voice agents for support and research.
• How to Implement Use ElevenLabs with per-language prompts. Start with low-risk queues. Add human-in-the-loop for escalation. Measure CSAT and AHT.
• Risks to be aware of Compliance and privacy. Edge cases. Accent and cultural nuance. Add a fallback to human agents.

Noteworthy Observations and Unique Perspective

• Treat content and customer stories as launches. Quote: “Everything is a launch.”
• Edit prompts, not outputs, to scale quality. Quote: “Try and edit the underlying prompt rather than the actual output.”
• Marketers can and should code small systems. Quote: “I did the first 90% in one day in Cursor.”
• Local-first MCPs can turn personal feeds into strategic intel. Quote: “Summarize the thoughts on ElevenLabs from the messages.”

Companies, Tool and Entities Mentioned

• ElevenLabs
• Granola
• GPTs (OpenAI)
• Claude (Anthropic) + Claude Desktop
• MCP (Model Context Protocol)
• Cursor
• GitHub + GitHub Actions
• Payload CMS
• WhatsApp + Whatsmeow
• SQLite
• Zapier
• Salesforce
• Calendly
• Google Ads
• LinkedIn, X/Twitter
• Orkes Conductor (sponsor)
• Retool (sponsor)
• University of Texas Medical Branch, Amazon GenAI, Ramp (Retool customers)
• chess.com (Professor Wolf)
• Palantir, SpaceX (tone reference)

Linkedin Ideas

1. Title: How we turned case studies into a weekly launch engine

• Main Point: Use Granola + a custom GPT to ship case studies and tweets in hours, not weeks.
• Core Argument: Prompt governance and CRM triggers unlock consistent output.
• Quotes: “Everything is a launch.” “Edit the underlying prompt rather than the actual output.”

2. Title: We replaced a $40k SaaS and $100k agencies with a 200-line workflow

• Main Point: LLM translations via GitHub Actions beat legacy tools on cost, speed, and quality.
• Core Argument: Control over prompts is the quality lever. Build when vendors block it.
• Quotes: “This saved us $40,000… Over $100,000 in agency costs.”

3. Title: MCPs make your messages actionable: What my WhatsApp taught me this week

• Main Point: Local WhatsApp MCP + Claude = trend summaries, drafts, and voice roundups.
• Core Argument: Agentic workflows beat brittle automations for changing tasks.
• Quotes: “You can on the spot come up with these agents…”

4. Title: Prompt governance is brand governance

• Main Point: Tone, examples, and format rules in prompts scale brand voice.
• Core Argument: Update prompts, not outputs, to keep teams fast and consistent.
• Quotes: “Give it as much context as possible.”

5. Title: Voice agents are the next support team you will hire

• Main Point: Spin up multilingual voice agents to scale support and research.
• Core Argument: Voice improves engagement and reach. Start with low-risk queues.
• Quotes: “You can spin up an entire team of customer support agents who are fluent…”

Blog Ideas

1. Title: The AI CMO Playbook: Make Everything a Launch

• Main Point: From feature to full-funnel in hours using prompts, MCPs, and orchestration.
• Core Argument: Distribution velocity is the growth edge.
• Quotes: “Translate every feature into your entire launch process.”

2. Title: Build vs Buy in the LLM Era: When to Kill Your SaaS

• Main Point: A decision rubric using prompt control, cost, and speed.
• Core Argument: If prompt control is blocked, quality and iteration die.
• Quotes: “I would much rather bet on AI costs getting cheaper and the quality going up.”

3. Title: Case Studies at Scale: Granola + GPT + CRM Triggers

• Main Point: Step-by-step system to produce 5+ case studies a month.
• Core Argument: Automate intake, standardize prompts, and publish fast.
• Quotes: “You’ll be turning out five case studies a month in no time.”

4. Title: Turning WhatsApp into a Knowledge Engine with MCP

• Main Point: Set up a local WhatsApp MCP to mine trends and drive content.
• Core Argument: Local-first preserves privacy and unlocks tool chaining.
• Quotes: “It downloads all your messages… Then you can keep querying it.”

5. Title: Voice Modality: From Demo Flair to Operational Powerhouse

• Main Point: Use voice to prototype narratives and scale multilingual ops.
• Core Argument: Voice increases engagement and reach; pair with human review.
• Quotes: “Prototype keynotes in Studio… Then send audio for timing and flow checks.”