How I AIThe exact AI playbook (MCPs, GPTs, Granola) that saved ElevenLabs $100k+ & helps them ship daily
CHAPTERS
Prompt-first editing mindset + episode promise: “vibe marketing” that ships daily
The episode opens with Luke’s core principle: improve systems by editing the underlying prompt/workflow, not manually tweaking outputs. Claire frames the conversation as “vibe marketing”—using AI to turn everyday work into repeatable launch assets and real cost savings at ElevenLabs.
- •Edit the prompt/system rather than patching AI output by hand
- •AI as a leverage tool for growth teams, not just engineers
- •Teaser of key workflows: case studies, social content, translation automation, MCP/WhatsApp
- •Goal: make “everything a launch” through automation
The AI CMO: automate launches from feature → messaging → distribution
Luke describes why marketing becomes the bottleneck as software creation accelerates. He outlines an AI-powered launch checklist that converts new features into value props, messaging, assets, and ongoing channel experimentation (ads, landing pages, optimization).
- •Software output is exploding; adoption/marketing becomes the constraint
- •ElevenLabs uses a detailed launch checklist for every feature/model
- •AI CMO vision: auto-generate blog/X/creative assets plus evergreen growth loops
- •Beyond launch: keyword discovery, Google Ads creation, landing page optimization
Live case study workflow: record a customer interview with Granola
Luke demonstrates a lightweight case-study interview using Granola to capture transcript and notes. The point is to reduce friction: short interviews produce structured inputs that can be immediately turned into publishable marketing content.
- •Use a fast interview format (a couple targeted questions)
- •Granola captures transcript + produces an auto-summary
- •Pull concrete facts, use cases, and metrics during the conversation
- •Goal: turn a few minutes of talk into reusable content assets
Custom GPT “Copy Editor”: enforce brand voice with rules + examples
Luke walks through an internal custom GPT used across the company to match ElevenLabs’ tone and style. It includes explicit role definition, strict must-do instructions, voice/tone guidelines, and reference examples (tweets and blog posts).
- •Custom GPT as a shareable, standardized prompt for the team
- •Identity + job-to-be-done (‘expert editor’ in ElevenLabs style)
- •Hard constraints: American English, serious research-led tone, word preferences
- •Use strong positive examples; discussion of when/where ‘bad examples’ help
From notes → publishable case study: use summary + raw transcript for fidelity
Luke shows how he pastes both Granola’s summary and the raw transcript into the custom GPT to generate a case study draft that’s usually usable on the first pass. Including the raw transcript preserves exact quotes and prevents “lossy” summarization issues.
- •Combine Granola summary (structure) with raw transcript (quotes/detail)
- •Skimmable headers designed as mini-summaries for readability
- •Iterate by asking for fixes (links, product details, pricing, SEO)
- •Key insight: context windows enable using both high-level and raw inputs
Turn the case study into distribution: tweet threads with asset placeholders
Luke demonstrates generating an X (Twitter) thread from the same source content, including guidance on what visuals to include. Claire highlights how this accelerates distribution by producing not just copy but a content packaging plan.
- •“Everything is a launch”: content creation plus distribution assets
- •Tweet thread structure: hook → bullets/short paragraphs → CTA/media
- •Model outputs include placeholders for screenshots/images to make posts effective
- •Extendable to LinkedIn posts or founder-voice rewrites via additional GPTs
Make it a system: Zapier + CRM triggers to produce case studies continuously
They zoom out from one-off creation to operationalizing it as a repeatable engine. Luke proposes connecting Salesforce closed-won events to automated outreach and scheduling, so the team consistently generates multiple case studies per month.
- •One-off efforts die when teams get busy—systems survive
- •Zapier flow: Closed-won → email + Calendly → interview → transcript → GPT output
- •Optional: GPT prepares interview topics ahead of the call
- •Outcome: consistent cadence (e.g., ~5 case studies/month) with minimal overhead
Prompt engineering best practices: tighten instructions and feed learnings back
Claire and Luke dissect why the prompting approach works: clear role, specific requirements, formatting constraints, content-type sections, and examples. Luke reinforces the practice of updating prompts whenever recurring issues appear rather than editing final text repeatedly.
- •Use explicit identity (“You are a…”) + concrete job definition
- •Write ‘must-do’ requirements and break down by output type (blog vs tweets)
- •Examples are underrated; bad examples can be useful in some workflows
- •Continuous improvement loop: fix the prompt, not each individual output
Replacing localization SaaS + agencies with an LLM translation system ($140k+ saved)
Luke details how ElevenLabs scrapped an expensive localization tool and agency-based translation pipeline due to cost, delays, and poor quality. He built a small server and prompt-per-language approach—mostly in Cursor—that delivers instant translations and reduced vendor overhead.
- •Old stack: $40k/yr tool + ~$100k agencies + days-long turnaround
- •Core failure: poor AI translation in tool + inconsistent agency quality
- •New stack: small service that sends strings to LLM with per-language prompts
- •Human spot-check only for sensitive pages; improve via prompt updates
Why build beats buy (sometimes): SaaS risks and ‘human-in-the-loop’ exposure
Claire and Luke discuss broader implications: when build cost drops, teams will increasingly replace niche SaaS—especially where value is routing low-skill human labor. Luke argues SaaS isn’t dead, but products dependent on manual labor arbitrage are vulnerable as AI improves.
- •Existential SaaS pressure: building custom tools becomes cheaper/faster
- •Marketers can now prototype solutions in Cursor and hand off to engineers
- •Human-in-the-loop models are risky where AI quality is rapidly improving
- •AI costs likely drop while translation quality rises—changing ROI calculus
Open-sourcing the translation approach: GitHub Actions + CMS ‘Translate’ button + Cursor rules
Luke explains the concrete implementation and why it worked better: prompt control, one source of truth, and automation at the code/CI layer. The system translates strings on each push and adds a CMS translation button, plus a Cursor rule to extract strings into JSON cleanly.
- •Key differentiator: ability to edit prompts (tool vendor lacked this)
- •GitHub Action: detect translation key changes → call LLM → write back JSON
- •CMS integration: a ‘translate’ button for content, maintaining consistency
- •Cursor rule automates string extraction/formatting (SSR/CSR handled)
MCP explained via WhatsApp: give AI agents tool access to your messages
Luke introduces Model Context Protocol (Anthropic) and why it matters: it lets agents use external tools and data sources. He built an unofficial WhatsApp MCP to download messages locally into SQLite and let Claude query/summarize them or send messages/voice notes.
- •MCP = protocol to expose tools/data sources to AI agents
- •WhatsApp pain: too many groups/messages; AI can’t help without context
- •Bridge emulates WhatsApp Web (barcode scan) using WhatsMeow; stores locally in SQLite
- •MCP server exposes query + send-message/voice-note capabilities to Claude
From rigid automations to flexible agent workflows: stitching MCPs together
They contrast static automation (Zapier/n8n) with chat-driven, tool-using agents that can adapt to new tasks on the fly. Luke shows combining multiple MCPs (WhatsApp + ElevenLabs) to create workflows like summaries and even voice roundups, pointing toward “tabs going away.”
- •Rigid workflows break when tasks change; agents can improvise with tools
- •Claude uses natural language to choose from exposed tool endpoints
- •Combine MCPs: pull WhatsApp insights → format in brand GPT → generate audio via ElevenLabs
- •Thesis: centralized chat interface reduces context switching and replaces ‘tabs’
On-demand specialized agents + lightning round: voice unlocks new products and ops
Luke demonstrates creating specialized agents on the fly (e.g., a case-study interview agent; mentions pizza-ordering demos). In the closing lightning round, he argues voice enables new customer experiences (tutors/coaches) and scalable back-office functions like multilingual support.
- •Create task-specific agents quickly by generating prompts/agent configs
- •Early days: many MCP projects are ‘toys’ now but signal future direction
- •Voice modality: enables interactive tutors (education, games, learning)
- •Operational unlock: multilingual customer support/research without hiring per language
