Acquired ACQ2 — Tobi Lütke (Shopify) on living in everyone else’s future

Why this is in the vault

This is the highest-signal AI/founder-mode interview the vault has captured to date. Tobi Lütke is one of the rare public-company CEOs who codes daily, is in the AI tooling weeds personally, and has internal practices that map almost 1:1 onto the harness/skills/evals language the vault has been formalizing throughout April. The conversation isn’t about Shopify the business — it’s about Tobi’s mental model for how to operate as a knowledge worker (and as a CEO) inside a phase change in computing capability. In the vault for three reasons:

He validates the “harness era” thesis from outside the AI labs. Cobus Greyling, Harrison Chase, Anthropic’s Thariq, and Karpathy have all separately argued that the locus of leverage has shifted from model weights to the surrounding scaffold (evals, context, skills, constitutions). Tobi describes the same thing in operator language — “context engineering,” “constitutions,” “running Toby evals on every new model” — without using any of the AI-lab vocabulary. Independent convergence is the strongest evidence the frame is real.
The “fall in love with problems, not solutions” frame is the cleanest articulation of why specialists are exposed and generalists aren’t. Tobi’s argument: the people who grieve when AI eats their solution are the ones who fell in love with the form of their work (the title, the role, the technique), not the problem the work was solving. The people who fall in love with problems are the ones AI most enables. This is a sharper version of the standard “AI doesn’t replace you, AI augments you” line — and it has direct implications for how RDCO writes about career adaptation, hiring, and the Sanity Check audience.
The “constitutions” practice — write down your uniqueness so AI can use it — is exactly the discipline the vault has been pushing toward. Tobi describes Shopify’s internal product principles as a living document hill-climbed against AI. Any project review can be tested against the constitution. The constitution gets sharper because AI flags imprecision and contradiction. This is a perfect parallel to what the RDCO vault is supposed to be — a knowledge graph of decisions, principles, and evidence that Claude can run any new piece of work against. Tobi’s been doing it at Shopify scale for months.

Core argument

The phase change is real and most people aren’t internalizing it. Tobi’s framing: there is software that no one wrote, and his job is to interview it for capabilities. His time allocation has shifted to (a) testing each new model against a personal eval suite, (b) figuring out where the edges of the new capability are, (c) figuring out how to make the models “idealized non-judgmental teachers.” He calls it “a privilege of a lifetime to be part of another platform shift” and he’s actively choreographing his attention around staying ahead of it.
He runs personal evals on every model — has for years. “Toby evals” = a folder of prompts with expected and judged results. He runs every new model against it and decides whether to switch. He maps this to “unit tests,” and the hosts agree. The point is not the eval suite specifically — it’s that the discipline of “test every new capability against a personal benchmark” is now table stakes for any knowledge worker who wants to stay ahead.
“Context engineering” > prompt engineering. Tobi defines context engineering as the skill of “stating a problem with enough context such that without any additional information the task is plausibly solvable.” This is harder than prompt engineering because it requires you to know what’s missing from your own statement. Side effect: he writes much better emails, runs much better meetings, and finds that “much of what people describe in companies as politics is actually bad context engineering.” Most disagreement at root is unstated assumption divergence.
Constitutions: write down your uniqueness, then hill-climb it with AI. Shopify uses Anthropic’s term “constitutions” for documents that capture “all the things where someone else would plausibly take the other side.” If no one would take the other side, it’s a platitude — drop it. Examples: Shopify product principles, area-specific principles. Project reviews now run against the constitution. The document gets longer and sharper over time because AI flags contradictions and imprecision. The discipline is to write down trade-offs you choose, not values you espouse. This is the closest thing to a thesis on how to make organizational tacit knowledge explicit and AI-actionable.
Use AI as the first pass on everything. Tobi sent a memo making this a Shopify-wide expectation. Reasoning: the people who use AI reflexively will compound advantage; not using it is unfair to your future self and to your team. Even if AI is bad at the task, you’ve now built an eval for the next model. The memo got leaked, was controversial, and he says it “now seems obvious.”
AI as your idealized self. Tobi’s most striking practical observation: he started asking AI what his next question should be, and “I get better ones very often.” The frame is that AI executes against an idealized version of you — the version of you when you’re at your best, not your average. This generalizes: AI is most useful when it raises the floor of your worst output to be closer to your best output, not when it tries to exceed your best.
“Concrete > abstract” emotional asymmetry. When Acquired ships an episode and millions hit the dashboard, Tobi notes (and Ben confirms) the emotional response is less than when they do a 1,000-person live event at Radio City. Our brains are wired for the concrete. This has implications for product, for marketing, for talent management, for how leaders communicate. Tobi treats this as a fundamental hardware constraint that won’t be solved.
15-year keylogger archive. Tobi has run a keylogger and screenshot-every-10-minutes of his own machine for 15 years. He didn’t think it would matter. Now he’s processing it with AI on weekends. “How Tobi became Tobi. How Shopify became Shopify.” This is the most extreme personal-archive case in the vault and it points to a broader truth: the people who get the most out of AI in 2026+ will be the ones who’ve been logging their own work all along. Personal data lakes become personal alpha.
The “fall in love with problems” career frame. Job titles and roles are calcifications of how a particular company solved a particular problem at a particular moment. The problem was the real thing; the role is the artifact. “You’re not a sales operations expert, you’re an amazing human being hired into a company and the company wants you to add as much of what you’ve got to the mission.” People who orient around the problem (not the solution) are the ones AI most enables — because the solution they were trained on is now cheap, and the problem definition is the new scarce skill.
Founder-mode through the AI transition: pick a thing you care about and bridge from a future to people you care about. Tobi’s prescription for anyone trying to find their footing in this moment. He builds Shopify; that’s his bridge. The transferable advice: don’t try to be everywhere; pick the one bridge you can build credibly and keep building it.

Mapping against RDCO

Direct validation of the harness/skills/evals architecture the vault has been building. Tobi independently arrived at: (1) personal eval suites per model, (2) context engineering as the core skill, (3) constitutions as living documents, (4) AI-first as a workflow expectation. The RDCO vault, ~/.claude/skills, the audit log, the graph, the knowledge graph schema — all of these are exactly what Tobi describes Shopify doing internally. The thesis is no longer speculative; the largest non-AI-native software company is doing it. This is the citation when a Sanity Check piece needs to argue “this isn’t a hype frame, the people building production software at scale are organized this way now.”
“Constitutions” → vault concept page candidate. Most RDCO writing has principles embedded in prose. The Tobi/Anthropic frame is to extract principles into a living document, hill-climb them against AI, surface contradictions. The RDCO vault should have an explicit ~/rdco-vault/02-strategy/constitutions/ directory: editorial constitution (Sanity Check voice and trade-offs), advisory constitution (RDCO consulting trade-offs), Squarely product constitution. Each project review and each newsletter draft can be tested against the relevant constitution. This is the natural next step after the existing draft-review and voice-match skills.
“Context engineering > prompt engineering” should be the canonical framing in any RDCO content about working with LLMs. Replace “prompt engineering” language wherever it appears. The skill is stating a problem with enough context that no clarification is needed — and learning to detect what’s missing in your own statement. This is also a useful management frame: “much of what looks like politics is bad context engineering.”
“Fall in love with problems, not solutions” is the right Sanity Check angle on AI career displacement. The vault has multiple drafts and reference notes about AI’s effect on knowledge work. The Tobi framing — your role is calcified solution, the problem was the real thing — is sharper than anything else in the vault on this topic. Worth pulling into a future Sanity Check piece directly. (Note: founder declined the Sanity Check angle suggestion in cycle 5 — keep this for the editorial calendar pool, not as an immediate brief.)
Personal data lake / 15-year keylogger as an extreme case for “log everything you do, AI will use it later.” RDCO’s vault is the institutional version of Tobi’s personal archive. The pattern: capture aggressively, structure later, query with AI. The implication for the founder personally: if there’s a daily-journal or activity-log practice that isn’t being captured now, the marginal cost of starting is approximately zero and the future value compounds.
“Idealized self” as the right metaphor for what good AI use looks like. Better than “augmentation” or “copilot.” The frame is that AI raises the floor of your output to closer to your peak — not that AI exceeds your peak. This implies a specific quality discipline: the human still has to know what their peak looks like, and the AI is most valuable when used to close the gap from average-output to peak-output. This frame should inform how RDCO talks about its own use of Claude in the operations stack.
“Concrete > abstract” emotional asymmetry has direct implications for Sanity Check distribution. The dashboard tells you nothing. Live events, in-person meetings, founder-replied DMs — those create the emotional traction that builds the brand. Even at 5x scale, the dashboard will still feel like nothing, and the in-person moments will still feel disproportionately big. This argues for keeping a meaningful in-person/concrete component to RDCO’s go-to-market regardless of how big the digital reach gets.

Open follow-ups

Constitution directory in the vault. Create ~/rdco-vault/02-strategy/constitutions/ with three living documents (editorial, advisory, Squarely product). Each should capture only trade-offs the founder would defend against the opposite position — drop platitudes. Run draft-review and voice-match skills against the relevant constitution. Concrete vault contribution.
“Context engineering” concept page. Define the discipline, contrast with prompt engineering, give 3-5 examples from RDCO’s own work (a good Notion task, a good handoff doc, a good vault article frontmatter). Cross-link to existing drafts.
“Fall in love with problems” concept page. Tobi’s frame, plus parallel quotes from the vault (probably Patrick Collison, probably some Joe Reis material). Cross-link to career-moats series in the commoncog references. Concept article candidate.
“AI as idealized self” frame. Worth writing up as a vault concept distinct from “augmentation” or “copilot” — those frames are too generic. Tobi’s specific articulation deserves its own reference.
Curiosity question: are there other non-AI-lab CEOs publicly running personal eval suites? Tobi is the example. Who else? Sundar? Dario? Zuck? Karpathy obviously, but he’s an AI-lab person. The interesting names are operators. Worth a research-backlog candidate.
Tobi’s keylogger archive is a pattern to study. The principle “capture now, structure later” applies to every domain of knowledge work. Any vault concept on personal knowledge management should reference this case as the extreme example.

Sponsorship

This episode included sponsor reads from:

Plaid — Mid-roll. Long-form sponsor read from the hosts about Plaid’s role in fintech infrastructure (Robinhood, Venmo, Chime). Hosts also reference their own ACQ2 episode with Plaid CEO Zach Perret. Treat as paid placement; not a subject of analysis in this episode.

This is a single-sponsor ACQ2 episode. The Tobi conversation is editorial and unpaid. Note: ACQ2 episodes typically run lighter sponsor loads than the main Acquired feed; this is normal for the format.