In April 2025, Shopify CEO Tobi Lütke posted an internal memo titled “Reflexive AI usage is now a baseline expectation at Shopify”. It went so viral that the AI memo became a genre; Box, Fiverr, and half of tech leadership published their own within weeks.
A year later, the mechanics behind the memo came out: Farhan Thawar, Shopify’s head of engineering, walked The Pragmatic Engineer through the whole system, and Bessemer wrote it up as an AI-first engineering playbook. Roughly 3,000 engineers, and productivity up about 20%, a number Thawar himself calls a humble estimate.
Everyone read it as an engineering document. Read it again through a marketing lens and it becomes something else: an operating manual for an AI-first marketing team. The translation is nearly one to one.
The core shift: produce becomes orchestrate
The playbook’s premise is that engineering’s job is moving from writing code to directing agents and evaluating their output. Bessemer summarizes Thawar’s guiding question as: “How do I get more of the AI to focus on the toilsome parts of writing code so that I can focus on the strategic parts?” Senior engineers at Shopify launch several agents at once, review what comes back, discard what fails, and merge what works.
Marketing’s version of the shift is identical. You stop writing the post, making the ad, and building the email.
You set the strategy and the taste bar, then review, merge the good, and bin the rest. The same review-merge-discard loop, run over copy and creative instead of pull requests.
What I’ve learned running this loop daily: agent output quality tracks context quality, and the curve is steep. An agent with your style guides, your product facts, a memory of past campaign learnings, and MCP access to your actual tools produces work at a completely different level than a chat window with a clever prompt.
Proper context and tools take output from “mediocre, needs a rewrite” to “we can ship that.” Most of the five principles below are really about building that context.
1. Standardize infrastructure, not tools
This is the strongest parallel in the whole playbook, and the one marketing has thought about least.
Shopify didn’t pick a winning AI tool and roll it out. They built the layer underneath: an LLM proxy that routes every AI request through one gateway (born as a privacy measure, Thawar explains in the talk, so employees stopped pasting customer data into public chatbots), token spend visible per team and per person, a leaderboard that celebrates the biggest token spenders, and MCP servers on everything, including the internal wiki.
Thawar’s phrase in the talk is that Shopify are “fans of MCPing all the things.” Tools on top stay swappable; engineers use Claude Code, Cursor, Copilot, and Codex side by side, because nobody knows yet which tool wins.
The marketing equivalent has two halves.
The first half is a source of truth for your foundational marketing content: product taglines, positioning, style guides, tone and voice, and skills that document your actual workflows. At GravityKit this lives in a shared repository the whole team connects to through Claude Code, versioned through GitHub.
When the voice guide improves, every agent and every teammate gets the improvement on their next session. That’s the marketing version of Shopify’s prompt library, and it’s buildable in an afternoon.
The second half is harder: the data layer. Marketing runs on cross-functional data (support tickets, analytics, ecommerce metrics, revenue, the engineering roadmap, the CRM), and an agent that can’t see that data can only produce generic work.
The first step is MCP connections to each source; my own setup has around twenty, covering analytics, the BI database, the support desk, the CMS, keyword research, and ads. The smarter long-term move is a centralized warehouse the agents query directly, which is where the sources stop disagreeing with each other.
Most marketing teams have neither half. The current state of AI adoption in marketing is a ChatGPT tab open in everyone’s browser, which amounts to standardizing on a tool (the thing Shopify deliberately avoided) while building none of the infrastructure. The playbook says to do the opposite, and on this point I think it’s simply correct.
2. Run agents in parallel, with critique loops
Bessemer frames 2026 as the year of orchestration, quoting Thawar: “If you don’t figure out how to harness agents in 2026, you’ll be behind.” The two patterns named in the write-up are parallel execution (several agents at once, human reviews and merges) and sequential critique loops (a piece of work refined through rounds of model self-interrogation).
The translation here needs adapting rather than copying. Engineering has long-running, deterministic tasks with a test suite as the referee, so agents can review each other’s work with little supervision. Marketing mostly doesn’t.
My rule after two years: oversee multiple agents working on separate tasks you can see, rather than delegating to sub-agents with no visibility. The exception is research, where fan-out works beautifully; my competitor comparison pipeline sends parallel research agents out and then runs adversarial fact-checking agents against their findings before a word gets drafted.
That adversarial pass is where critique loops earn their cost in marketing: fact-checking and second passes on data analysis. A second model interrogating “did this number actually come from the source” catches real errors. Critique loops on matters of taste mostly produce confident agreement.
The unglamorous work that makes any of this function is routing: each task type needs a defined path to the right reference files, memories, and data stores. I’ve written before about how that setup works in practice.
3. Declare a Code Red on your content debt
My favorite story from the talk: Shopify spent seven months (they planned three) with 30 to 50% of engineering paying down tech debt. The debt was throwing off signals, growing exception counts and seg faults, and Thawar refused to keep shipping features on top of it. They set exit metrics instead of vibes: seg faults at zero, exception counts falling.
Marketing accumulates the same debt in its content. Stale posts, decaying rankings, orphaned pages, outdated screenshots, docs describing a UI that no longer exists. It throws off signals too: declining organic traffic, sliding click-through rates.
Most teams keep shipping new campaigns on top of it anyway, and the playbook’s answer is to stop, freeze the calendar, and pay it down. “Content debt” is my term, and the Code Red framing is my translation, but the discipline transfers exactly, including the part nobody wants to hear about freezing new work.
The stakes are higher now than when stale content merely ranked poorly. AI assistants read your content to represent your brand, so an outdated post recommending a feature that no longer exists is actively misleading the systems your customers now ask for recommendations.
Two things follow. You need a running system for keeping content current, an audit-and-refresh loop rather than a one-off cleanup. And you should give customers decision-making surfaces you control (your own chatbot grounded in current docs, or a free consult call) so a third-party AI’s summary isn’t the only path to a decision.
4. Right tool for the right layer
Thawar tells a story in the talk about trying to build a small product-availability checker in Claude Code. An hour in, it was still failing, because the coding agent was attacking the problem at the wrong layer.
Rebuilt in Gumloop, a browser automation tool that already knows how to navigate websites, it took two minutes. His conclusion: different tools for different jobs.
For marketing I’d sharpen that. The layer problem is real, but the fix is rarely swapping models; what matters is the tools the model can call. Same model, different tools, wildly different output.
My clearest example: I asked Claude to build a landing page through a popular WordPress MCP plugin, and the output was garbled. Pages didn’t flow and ignored our design system entirely.
The fix was building a custom MCP server that translates our pages into a structured format and gives the agent access to our component library. Same model on the same task now assembles pages from our actual design components, and they look right the first time. The bottleneck was the tool the model was holding.
So the translated principle: before blaming the model or writing a longer prompt, ask what layer the task lives at and whether the agent holds a tool that operates there natively. Sometimes that tool exists already (keyword data platforms, browser automation, image generation). When it doesn’t, you build it, and building it is exactly the infrastructure investment from principle 1.
5. Measure reversion, not output
Engineering already learned that output metrics are gamed the moment they become KPIs. Bessemer quotes Thawar: “Code is cheap now. But I don’t want code, I want solutions.”
Shopify tracks reversion rate, how often merged work has to be rolled back, and reports it has stayed flat while AI lifted PR volume. Nothing merges without senior human review. The honest progress signal, in Thawar’s view, is weekly demos.
Marketing’s output metric is just as cheap now, and just as gameable. “We published 40 posts this quarter with AI” is the lines-of-code of marketing.
The signal worth watching is how much AI output needs heavy rewriting before it ships, and how much gets binned, with conversion and engagement as the merge signal. Taste is the human guardrail, exactly as senior review is Shopify’s.
Cheap output also raises the bar on what deserves publishing at all. If a piece can be synthesized by AI from existing sources, your reader’s assistant will do that synthesis without you, so don’t write it. What survives is work with something novel in it: original data studies, or the guide nobody has written because writing it takes expertise.
On measurement itself, attribution keeps getting harder as AI referrals and dark traffic grow, and Rand Fishkin at SparkToro has been making this case for years. The practical fallback is a dashboard that pulls your key metrics and overlays publishing and campaign timelines, then reads correlations honestly instead of demanding a clean attribution chain that no longer exists.
Sharpen the axe
There’s a line attributed to Lincoln: give me five hours to fell a tree and I’ll spend the first four sharpening the axe. The context and the tools are the axe. Every principle above is sharpening; the drafting is the short part at the end.
Where to start: pick your most repetitive workflow, write down how you actually do it, and give an agent the files and tool access to do it your way. That’s the first piece of your infrastructure.
Shopify called it an engineering playbook. Read it as a marketer and it’s yours.