How Paul Rowsthorn Runs Airguides With an AI Co-Founder Named Amelia

Nathan Clark runs three clients on three different AI stacks. One has standardised on Claude, the next on ChatGPT, and he doesn't have the luxury of picking just one. So he built a system that lets him move between models and agent apps without starting over each time. Nathan Clark, founder of ReadyGTM, helps companies sharpen and scale their go-to-market motion, with a growing focus on AI implementation. On screen, he opens the file system underneath his work and hands a live piece of client work from one tool to another, showing exactly where the handoff still gets messy. The system is a folder structure on an always-on Mac Studio, reachable from his laptop, iPad and phone through Tailscale. The folders follow how Nathan runs his business: plan, market, sell, deliver, operate. Each client gets an AGENTS.md file, and a job that runs at 10pm each night sweeps his messages and files to keep those client files current, but it's only allowed to touch low-risk files. A centralised skills library is symlinked into every tool he uses, so the same skills are available whichever agent he's in. For the demo, Nathan works from a synthetic client, Acme Inc. He opens a sales diagnostic built in Codex, his daily driver, and explains why he delivers client reports as web pages rather than slide decks: clients are paying for clarity, and the evidence behind every number sits one click away. When he doesn't like parts of the design, he hands the same folder to Claude Code in the terminal and dictates the brief: strip the eyebrows and page numbers, add a dark-mode toggle. While that runs, he shows GrokBot answering client questions from the same underlying files. The change lands, dark mode works, and he goes back to Codex to push it live. The back half covers how Nathan picks tools. Codex does around 80% of his paid work because it follows precise instructions literally. Claude handles visual work and reads intent from looser prompts. Solo runs multi-agent builds for hours while he's elsewhere, and Amp's cloud-based Orbs keep environment and context intact without anything running on his desktop. Then the rough edges: memory drift in his file system, and agents that have pushed changes live when told not to. Nathan wants proper per-agent permissions rather than instructions an agent may or may not follow. It closes with a quick-fire round and Nathan's advice for anyone starting out: be more ambitious about what you ask AI to do, and when it names something you don't know, ask what it is and ask it to walk you through doing it. As he puts it, it's prompts all the way down. Nathan Clark is the founder of ReadyGTM, where he helps companies sharpen and scale their go-to-market motion and put AI to work in ways that move the numbers. He also works with Works.

Román Olivera runs his company on files and folders. No database, no CRM platform — a fractal system he calls Flatfish, built from plain text and a set of rules an agent reads before it does anything else. On this episode of Operator AI, he opens it up and shows exactly how it works. Román Olivera, from Orban Labs, walks Zachary King through the system live. Every folder that matters is a "node," and every node carries the same four things: rules for how to behave there, a current-state file for what's true now, an append-only log for what happened, and a registry so an agent can find its way around without being briefed from scratch. Román built it after losing track of one too many half-finished projects scattered across folders he couldn't remember the naming convention for. The demo is the receipt. By voice, from his phone, Román creates a CRM entry for a contact met at the local pub, in Codex running on a remote machine. Codex then builds an HTML presentation in Román's own design system. Then Claude Code, a different tool running a different model, picks up the same context in a fresh harness. It uploads the deck to AWS S3, shortens the link, and drafts two follow-up emails: one to Zac in English, one to the new contact in Spanish. Nobody told it to switch languages. It read the contact's file, which noted where he was from, and made the call itself. That portability is deliberate. Román turns platform memory off on his agents, on purpose, because an agent with its own memory tends to skip reading the rules and shortcut toward an answer. He'd rather own the context outright than rent it from whichever platform is trending: he uses APIs over MCPs so he can control the process, and keeps his skills in a plain folder rather than a config file, so nothing locks him into one tool. The conversation lands on an analogy Román coins mid-episode, on the record, timestamped live by Zac: AI is bringing back the Mad Men era of knowledge work, where the job is to direct rather than to do. The advertising executive never touched the filing cabinet — he told someone where a letter should go and trusted it would be there when he asked for it. Román argues that's the discipline AI now demands: be specific about what you want and where things belong, and the system holds up the rest. Román also draws a line most operators haven't hit yet. Files and folders work brilliantly for a business of one, or close to it. Once client work moves into enterprise environments, the same logic has to move into databases — a challenge he's already started tackling with clients on platforms like Azure. Before Orban Labs, Román spent more than a decade in the energy sector across field engineering, project delivery, and operational leadership — work that shaped his insistence that speed only matters once a system is reliable enough to trust with real work.

Luke Marshall, CEO at GradEngage runs growth marketing from a single Claude Code instance — hosted on a cloud VM so he can reach it from any device, tied to a running knowledge archive, and extended into everything from lead generation to his own workout tracker. GradEngage helps companies keep graduate hires warm and prepared in the months between accepting an offer and actually starting the job. Luke, who describes himself as the growth marketing automation lead among his co-founding team, opens his screen to show the system behind it: a landing page and product video rebuilt with Claude Code and shipped to WordPress, and a lead-generation workflow that scrapes public prospect data to auto-generate personalised landing pages and email previews for each target company. The setup starts with infrastructure. Luke hosts Claude Code on a cloud VM synced to Dropbox, so it's reachable from any device, and pairs it with a skill that finds his latest reference notes without him re-explaining context each time. From there, he walks through a standalone Convex database built to track GradEngage's roughly 6,500-company target market — kept deliberately separate from the CRM so early-stage prospects can be nurtured slowly, without cluttering the pipeline tools the sales team relies on day to day. Judgement still sits with Luke throughout. He explains why he switches between typing and voice dictation depending on the task, and why he pushed back on a cold pitch offering to run his marketing on full autopilot. That judgement call gets tested directly in the episode's most useful failure: an outreach automation tool re-messaged the same contacts with no response, burning leads before Luke caught it. The fix wasn't more automation — it was less. AI now runs read-only on outbound messaging, tracking open threads for a human to action rather than sending anything itself. The system extends past work, too. Luke shows a custom workout generator that pushes Claude Code output to his phone and watch, and a from-scratch AI assistant he built for his wife to handle weather, tasks, and a shopping list — settling on a cheaper model to keep it running affordably day to day. Luke Marshall works in growth and marketing automation at GradEngage.