Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
AI Agents run my business and life
TL;DR
Andrew Wilkinson is already running slices of a real business with AI agents — in his Deep Personality startup, support agents fix tickets and merge PRs, while a marketing agent hooked into PostHog, Meta, and Reddit manages creative, tests, and budgets on a product that has already done about $20,000 in revenue.
The biggest unlock for him is not one model, but a personal data layer — he records meetings with Fireflies, ingests email into GBrain, stores business data in vector databases, and gives agents searchable context across his family office, portfolio, health, and daily life.
He thinks 'fully autonomous companies' are mostly hype today — his analogy is that current agents are like Zapier zaps with judgment or a 'genius baby' that still needs step-by-step instruction, though he believes basic businesses could be far more autonomous within 3 to 6 months.
AI has turned non-technical operators into internal software builders — Wilkinson’s CFO, who had never coded before, replaced a $50,000 to $100,000-per-year Addepar-style wealth dashboard with a custom in-house portfolio system built in roughly two weeks using Claude Code.
His broader market take is blunt: software moats are collapsing fast — as vibe coding makes app creation cheap and fast, he expects pricing pressure to intensify across SaaS, compares many once-great software businesses to Buffett-style 'cigar butts,' and says he’s personally leaning toward hard-to-replicate assets and even data-center bets like TSMC.
His most practical prompting advice is to make the model interview you first — instead of trying to write the perfect prompt, he tells Claude his goal, asks it to 'ask me a shitload of questions,' and often has it use teams of eight sub-agents to produce much stronger outputs.
The Breakdown
From curiosity to full-blown Claude Code obsession
Wilkinson opens by saying December 2025 was the moment everything flipped: tools had been interesting before, but then Claude Code hit and he started waking up at 3 or 4 a.m. just to sit in terminal with ten tabs open. He jokes that the experience felt like “shooting heroin,” and that he’s been chasing that first peak ever since.
The Arizona moment that sold him on agents
The big proof point came on a trip to Arizona, where he forgot his laptop but had an OpenClaw agent running back home. From the back of Ubers, he ran his business through the agent, had it write all his emails, and says nobody noticed — which gave him that “chasing the dragon” feeling. The catch: he now spends a huge chunk of his time debugging and tuning the system instead of just benefiting from it.
Turning a relationship experiment into Deep Personality
What started as a project for him and his girlfriend became a startup after they fed psychological test results into ChatGPT and got a relationship analysis that “nailed every single fight.” He used Claude Code to convert 15 different screens into JSON-based multiple-choice tests, then built Deep Personality, which generates a long, Robert Greene-style report covering archetypes, attachment, ADHD, OCD, job fit, and relationship dynamics. He loves that vibe coding let him ship the whole thing himself instead of managing a chain of designers, engineers, and copywriters.
Harbor, support bots, and a marketing agent buying ads
To run the business, he moved from OpenClaw to Harbor, a GUI-based agent harness built by his friend Gavin Vickery. Inside it, support handles tickets automatically and can escalate bugs to a dev agent that generates fixes and PRs; for P0 issues, it can even merge the fix itself. The marketing agent is even wilder: tied into PostHog, Meta, and Reddit, it creates ad creative, runs multivariate tests, and adjusts spend — and Wilkinson is now wondering what happens when he hands it a $100,000 monthly ad budget.
Why he thinks autonomous-company startups are overselling it
When Greg brings up companies pitching “autonomous companies,” Wilkinson pushes back hard. His view is that today’s agents are more like smarter Zapier flows than executives: useful, but needy, like a “genius baby” or intern who still has to be told exactly how to think and what to check. He does think Anthropic or OpenAI will eventually ship something closer to an AI CEO, but says we’re not there yet because memory and context are still too fragile.
Building a searchable brain for his family office and conglomerate
A huge part of his stack is data plumbing: Fireflies records meetings, nightly cron jobs turn them into markdown, and vector databases make giant datasets queryable by LLMs. He demos Foley Partners data showing 132 direct investments, $16 million invested, and a current value of $36 million, then explains how he uses similar systems inside Tiny to ask for “icebergs” in quarterly performance or simple operational questions across 24 businesses. It’s not perfectly accurate, but he says it gives him an “eye of Sauron” inside companies that are otherwise too complex for one person to track.
The bearish software thesis and the $40,000 Claude bill
Wilkinson’s macro view is pretty stark: software is becoming easier to make, easier to copy, and harder to defend. He says his family office has effectively replaced headcount growth with API spend — “instead of a payroll, we just have a $40,000 a month Claude bill” — and gives the example of his CFO building a custom Addepar alternative in two weeks despite never having coded before. His conclusion is that many SaaS businesses are heading toward brutal pricing pressure, which is why he’s increasingly drawn to harder-moat assets and even infrastructure plays like Iron and TSMC.
Personal assistants for email, media, and health
The last stretch gets very personal: Wilkinson shows an agent that reads his email and iMessages, identifies projects and priorities, and turns replies into multiple-choice options he can answer via Telegram with “1A” or “2B.” He also built a custom daily podcast from newsletters and Readwise using Gemini voice, plus a health agent that analyzes Apple Watch, Eight Sleep, and other data, correlating symptoms like nerve pain with changes in wrist temperature days in advance. He closes with his favorite prompt move: don’t write the perfect prompt yourself — tell Claude your goal, let it interrogate you for five to ten minutes, and have it spin up a team of eight sub-agents to do the work.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
