AI For Restaurants

Sterling and Aaron break down the new agent stack and what AI actually does to your org chart. Sterling argues the SMEs (subject matter experts) inherit the earth, while Aaron insists generalists who abstract across problems are equally superhuman now. They unpack the three reasons to spin up a sub-agent (parallelization, context preservation, bias-free devil's advocate), what makes a session "poisoned" and how to escape it, and why every Claude turn should produce a documented artifact and a git commit. Aaron reveals Thanx is burning close to a billion tokens a week — their monthly AI bill matches the cost of an employee, but instead of one new hire they've accelerated 75 people. Plus: slide decks are dead (Sterling's last board meeting was HTML, not PowerPoint), every engineer is secretly building their own PR-management tool, and Thanx's graphic designer is now shipping pull requests before the engineers finish reviewing them. Subject matter experts become Smees become superhuman — just don't call them Smeeze (Aaron only thinks of the fat pirate from Peter Pan). If you heard anything today that didn't make sense, just "ask Claude."

Sterling declares slide decks dead — his last board meeting was the last one he'll ever make. He's building HTML presentations with AI now and says in a couple months everyone else will too. Aaron pushes back with the Lawrence Lessig TED Talk approach: slides should be visual storytelling, not six bullet points read aloud. He's trained Claude on his own doodle style and now generates bespoke sketch-art slides that look like something he'd actually draw. They get into where practitioners actually learn about AI when LinkedIn is drowning in "comment CLAUDE for my free prompt" grifters. Aaron's method: find someone posting something brilliant, then go read the smartest comments on that post and follow those people — the record label approach to curating your feed. Sterling's honest: he just asks Claude. The memory vs skills debate gets heated. Sterling went from obsessed with memory (Neo4j graphs, Synapse analyzers finding 14 contradictions) to hating it and going all-in on skills. Aaron agrees skills are where the real training happens but argues memory has a purpose — it remembers the esoteric stuff that doesn't belong codified in a recipe. The cookie baking analogy lands: skills are recipes you iterate on, memory is knowing you prefer a certain brand of chocolate. Sterling's not buying it. If you heard anything today that didn't make sense, just "ask Claude."

Sterling brings the receipts on AI search. He's been tracking restaurant website traffic and watched AI referrals jump from 0.3% in 2024 to 2.5% in 2025, with a projection north of 10% in 2026 — roughly a 33x increase in two years. He and Aaron break down what it actually means: GEO (Generative Engine Optimization) is the new SEO, Google reviews are the new social proof for AI answers, and most of the "AI SEO" advice floating around LinkedIn is garbage. Sterling confesses he's on startup number six and has hit his Claude Max weekly cap four weeks in a row. Aaron's teaching an executive class where the first hour and a half is just 100 slides of "here's what's possible," and argues your employees are the dress rehearsal before any customer touches your AI. Plus: a new pair of skills — Sterling's /pillage and Aaron's /steal — that scrape other people's repos, articles, and setups and hand you a tiered list of exactly what to copy. Plus: Agent SDK's new session persistence and dispatch features, why form-fitting software is replacing lowest-common-denominator builds, and the classic director who reads a LinkedIn article and Slacks "hey guys, are we doing this?" Forget SEO — it's pillage and steal now. If you heard anything today that didn't make sense, just "ask Claude."

Sterling has 15 go-to-market agents running at Chowly right now. Not chatbots. Not "AI assistants." Autonomous agents with specialized context, dedicated tools, and triggers that fire on events — like an SEO agent that sees a demo get booked, pulls the prospect's website, runs a full analysis, and has the results ready before the rep even opens their laptop. He breaks down the architecture: context gives the agent its specialty, tools give it capabilities, and triggers give it autonomy. Without all three, it's just a fancy prompt.Aaron counters with a different flex: he got his entire company to go AI-first. Every department — sales, ops, finance, support, engineering — is building and shipping. In one week, 25 new skills and automations were created across the org. The ops team is now the biggest contributor, building RAG databases and customer inquiry agents. Salespeople are sharing repos in Slack. His definition of AI-first is precise: "Beginning every single task by collaborating with an agent to plan, allowing the agent to execute as much of that plan as possible, and then iterating on its skill until you can trust it to execute it well every time."The governance philosophy? Get everyone into motion before you add friction. Set red lines — no publishing to the internet, no third-party skills, no PII in the LLM, no API keys in Git — and let everything else burn tokens. Local apps graduate to hosted infrastructure only after technical review. And the secret weapon for onboarding? A bootstrap skill that points Claude at a getting-started guide and lets it walk new users through the entire setup.They close on context engineering — the journey from empty prompts to system instructions to memory files to skills — and why most people struggle with AI because they never teach it who they are. Sterling's co-founder put it best: "It's not about prompt engineering, it's about context engineering." If you heard something today that didn't make sense, just ask Claude.

Sterling and Aaron go head-to-head on OpenClaw. Sterling calls it a massive unlock, Aaron calls it a security nightmare waiting to happen. They break down why always-on AI assistants are so seductive, the original sin of computing that makes LLMs inherently risky, and why both of them use these tools anyway. Aaron reveals he's managing 30+ Claude instances at once and has turned his workflow into a session-parking machine that cranks through 10-15 plans a day. Plus: how Agent Teams in Claude Code are changing their workflows, why the "SaaS apocalypse" is overblown, and a confession that Claude Code has replaced video games as their late-night addiction. Forget Netflix and chill — it's Claude and chill now. If you heard anything today that didn't make sense. just "ask Claude".

Aaron Newton burns through close to a billion tokens a week. Sterling's head of HR, who has never written a line of code, built a training app that replaced their existing one, complete with user roles, tracking, and a GitHub repo. This is what happens when you stop asking AI for answers and start letting it cook.In Episode 2 of AI for Restaurants, Sterling and Aaron move past the "should I use AI" question and get into the "how". The conversation opens with both hosts comparing token usage like a scoreboard, and the numbers tell a story: the more you use AI, the more value you extract, and the cost is almost comically low compared to the output. Aaron describes paying roughly $70 a week for what amounts to a full software development team.From there, they dig into vibe coding and why it's both overhyped and genuinely powerful. Sterling's HR example is the proof point: a non-technical person built production software that the whole company now uses daily. Aaron shares how his head of marketing used AI to pull every competitive deal transcript from Gong, analyze what customers said, and build battle cards that actually reflect real sales conversations. The key insight in both cases is that the data had to be accessible first. If your call recordings, CRM data, and internal docs are scattered across platforms, AI can't help you. If they're organized, the results are immediate.The episode's standout segment is Aaron's explanation of Skills using the Matrix helicopter scene: Trinity needs to fly a helicopter, someone loads the knowledge, and she instantly can fly it. Skills work the same way for AI. They're text files written in plain English that teach your AI agent how to do specific jobs, from querying Salesforce to sending Slack messages. They're token efficient, easy to iterate on, and unlike MCPs, you don't need to understand Docker or authentication protocols to get started. Both hosts describe how skills have become the backbone of how they work.Sterling then walks through his personal AI task management system. He took a to-do app, embedded AI agents into it, and now when he adds a task, an agent picks it up, invokes the right skills, and either completes it autonomously or enters planning mode with him for anything complex. Aaron reframes it perfectly: "You took an app that manages your to-do list and turned it into an app that manages Claude's to-do list."The conversation wraps with both hosts discussing how they give their AI agents personalities based on real people they've worked with, and Aaron introduces Sherlock, his company's AI analyst that doesn't just answer questions but investigates them, chases down related threads, and flags its own assumptions through an adversarial review agent called Moriarty.

Every single person has a moment when they're using AI and they stop and go, "Holy SH**, am I out of a job?" That moment is the red pill. And how you respond to it will define your career, your company, and your ability to compete in an industry that's about to change forever.In the debut episode of AI for Restaurants, Sterling Douglass (co-founder and CEO of Chowly) and Aaron Newton (co-founder and CDO of Thanks) dig into what the "red pill moment" really means for restaurant leaders. This isn't a theoretical conversation. Both hosts run restaurant technology companies, use AI daily to operate their businesses, and have watched firsthand as the gap between AI adopters and everyone else has widened at an alarming pace.Aaron breaks down a hiring reality that's already here: AI-first managers will only hire AI-first employees. If you haven't started, you can't attract the talent that has, and you're locking your entire organization out of the game. Sterling shares his own red pill moment, going from copy-pasting ChatGPT into a terminal to running autonomous coding loops that execute plans while he sleeps.The conversation covers the practical mechanics of working with AI. Why planning mode changes everything. How MCPs connect AI to your real business data. What context windows are and why they matter. How "Ralph loops" can automate entire projects overnight. Aaron details his AI boot camps for restaurant CMOs, CEOs, and COOs, built around a simple premise: everybody programs, but no one writes code. Sterling explains why restaurants, with their deceptively complex business models combining showroom fronts and manufacturing plant backs, stand to gain the most from AI adoption.Both hosts are blunt about the stakes: the window to catch up is closing fast. In six months, "I'm curious about AI" won't be enough. Hiring managers will ask what you've actually built. Companies that drag their feet doom not just themselves but their employees to a future where they can't compete. This episode is a wake up call and a practical starting point for anyone in the restaurant industry who knows they need to move but hasn't yet.