How I AIThe exact AI playbook (MCPs, GPTs, Granola) that saved ElevenLabs $100k+ & helps them ship daily
CHAPTERS
- 0:00 – 2:44
Prompt-first editing philosophy + why these workflows save real money
The conversation opens with a key principle: improve the prompt/system, not the generated text, so outputs steadily get better over time. Claire tees up Luke’s track record of cutting tooling and agency costs using simple AI-enabled workflows.
- •Edit the underlying prompt rather than patching the output
- •Prompt structure: strong role + explicit requirements drives consistency
- •AI workflows can replace vendors/agencies and reduce operational overhead
- •Teaser of savings and automation themes that will recur
- 2:44 – 4:23
“Vibe marketing” and the AI CMO: turning every feature into a launch engine
Luke explains why the explosion of software creation makes distribution and go-to-market the real bottleneck. He describes an AI-assisted CMO future: automatically translating product updates into a full launch checklist and ongoing evergreen growth experiments.
- •Software output is accelerating; adoption and distribution become the constraint
- •ElevenLabs uses a structured launch checklist (value props → messaging → assets)
- •AI CMO vision: generate launch assets (blog, X, images/video) automatically
- •Beyond launch: spin up evergreen channels (ads, landing pages, optimization)
- 4:23 – 9:06
Live case study workflow: Granola interview → Custom GPT → publish-ready story
Luke demonstrates a fast path to producing case studies by recording a short customer interview in Granola and feeding the summary + transcript into a tailored ChatGPT GPT. Claire shares a concrete ElevenLabs use case, which becomes a polished draft in minutes.
- •Use Granola to capture the interview and generate a high-quality summary
- •Collect concrete use case details and at least one ROI/time-saved metric
- •Feed both summary and raw transcript to reduce lossiness and preserve quotes
- •Custom GPT outputs a structured, skimmable case study draft
- 9:06 – 11:52
Building an “ElevenLabs Copy Editor” GPT: tone of voice, examples, constraints
Luke walks through the exact components of their internal copy-editing GPT: role definition, brand voice guidance, and a library of strong examples. The goal is consistent writing that matches brand standards across blogs and social without heavy manual editing.
- •GPT as a shareable, reusable prompt artifact across the company
- •Embed tone-of-voice guide (serious, research-led; American English, etc.)
- •Add content-type-specific instructions (blog vs. tweets)
- •Include multiple high-quality examples to anchor style and structure
- 11:52 – 13:53
From case study to distribution: tweet threads, asset placeholders, founder voice variants
Luke shows how the same GPT can convert the case study into an X thread with media guidance, then be repurposed into other channel formats. The broader theme is treating content creation and distribution as one continuous “launch” workflow.
- •Generate tweet threads directly from the case study content
- •Use bracketed placeholders to specify required media assets/screenshots
- •Extend to LinkedIn posts and founder-voice rewrites via additional GPTs
- •“Everything is a launch”: creation + distribution designed together
- 13:53 – 15:15
Systematizing case studies with automation: Zapier + CRM triggers + repeatable pipeline
They move from a one-off demo to a scalable machine: automatically prompting interviews and content generation after a deal closes. Claire highlights the operational value: removing reliance on human memory and making asset production continuous.
- •Trigger outreach from Closed Won in Salesforce → email → Calendly booking
- •Optional AI pre-brief: suggested questions and talking points per customer
- •Run interview → Granola → GPT to produce multiple assets quickly
- •Automation keeps the content engine running despite shifting schedules
- 15:15 – 19:51
Prompt engineering best practices: specificity, formatting, and iterative prompt upgrades
Claire breaks down what Luke’s prompt does well—clear role, explicit must-do constraints, and examples—then asks about using negative examples. Luke reinforces the meta-lesson: fix recurring problems by updating the prompt, not hand-editing each output.
- •Effective prompts: identity/role + precise instructions + formatting rules
- •Separate guidance by output type (tweets vs. blog posts)
- •Mostly good examples for copy; negative examples appear later in translation
- •Iterate by folding feedback into the prompt for compounding improvements
- 19:51 – 23:04
Replacing expensive localization stacks: why the old tool + agencies failed
Luke explains ElevenLabs’ original localization approach—costly SaaS tooling plus agencies—and why it underperformed. The central failure: poor AI translation quality, inconsistent human output, and a system that didn’t allow prompt customization.
- •Need: localize web content across many languages at scale
- •Old stack: $40k/year localization tool + ~$100k agencies + engineering integration
- •Quality issues: AI translation ‘terrible’ and agency output inconsistent
- •Key blocker: vendor didn’t allow editing the translation prompt
- 23:04 – 25:14
Cursor-built translation server + prompts per language: $140k+ saved, faster turnaround
Luke describes the replacement: a lightweight service and GitHub-driven flow that sends strings to an LLM using a per-language prompt file, then writes translations back to the repo/CMS. It reduced costs dramatically and cut turnaround from days to instant with minimal human QA.
- •Built 90% in a day (as a marketer) using Cursor; engineer helped productionize
- •Savings: canceled $40k tool + eliminated $100k+ agency spend
- •Architecture: string in → per-language brand prompt → LLM → write back to GitHub/Payload
- •Human-in-the-loop only for sensitive pages; adjust prompts to fix issues
- 25:14 – 29:46
Open-sourcing the localization solution: GitHub Action, CMS “translate” button, Cursor rule
Luke says they intend to open source the workflow if there’s interest. He details key building blocks: a GitHub Action that runs on string changes, a CMS translate button, and a Cursor rule to extract and manage strings cleanly.
- •Plan: open source the solution based on community demand
- •GitHub Action auto-translates whenever translation keys change
- •CMS integration: translate button for blog/content localization
- •Cursor rule automates string extraction and JSON management
- 29:46 – 32:33
WhatsApp MCP explained: connecting private message context to AI (locally)
Luke introduces MCP (Model Context Protocol) and why WhatsApp is central to a useful personal assistant. He walks through an unofficial local setup that downloads messages to SQLite and exposes query/send actions to Claude via an MCP server.
- •MCP = protocol (Anthropic) to expose tools to AI agents
- •Why WhatsApp: high-signal personal + work context locked in chat threads
- •Bridge imitates WhatsApp Web login; downloads messages once to local SQLite
- •MCP server enables querying messages and sending texts/voice notes via Claude
- 32:33 – 37:20
Practical MCP use cases: summarization, trend mining, and chaining multiple MCP tools
They demo summarizing recent WhatsApp messages and extracting themes from busy group chats. Luke describes chaining MCPs (e.g., WhatsApp + ElevenLabs) to turn summaries into audio roundups or distribution-ready content.
- •Summarize recent messages and weekly group activity to reduce overload
- •Mine WhatsApp groups for tool chatter/trends and convert into tweet threads
- •Chain multiple MCPs: query → generate → send (or generate audio with ElevenLabs)
- •Local-first design reduces cloud exposure; supports ongoing message syncing
- 37:20 – 41:07
From rigid automation to on-demand agents: generating specialized agents in real time
Luke contrasts static workflows (Zapier/n8n) with flexible chat-based tool use that adapts to ad hoc tasks. They show how an AI can create another specialized conversational agent (e.g., for case study interviews) on the spot, hinting at the future of agentic systems.
- •Static automations break when tasks change (e.g., new research roundup request)
- •Chat-based MCP tool use supports dynamic, high-level task execution
- •Demo: create a conversational AI agent for case study interviews on demand
- •Still early/toy-like today, but trajectory points to broader agent orchestration
- 41:07 – 44:24
Lightning round: what voice unlocks + where to find Luke and his launch checklist
Luke answers how voice changes product possibilities: richer customer experiences (tutors/coaches) and scalable back-office functions like multilingual support. The episode closes with where to follow Luke and read his launch and growth-hiring playbooks.
- •Voice enables new customer experiences (education, interactive tutoring, coaching)
- •Operational leverage: multilingual customer support and data collection at scale
- •Luke’s resources: launch checklist post and growth hiring guide
- •Where to follow: harries.co and Twitter (Luke Harries underscore)
