AI for Marketing & Ops
I'm a marketer and a CPA, so I look at AI the way an operator does: does it move a number, and can I prove it? Most AI-for-marketing content is either vendor hype or generic prompt lists. This is the other thing. Here I track how AI actually changes marketing and operations work, replacing manual effort in real campaigns, SEO, and internal tooling, and where it quietly fails or adds cost. The test I keep coming back to: can a marketer ship a working tool before lunch, without a dev team? When the answer is yes, I show how. When it's no, I say so.
25 posts
- GaP treats robot policies as editable graphs A new robotics paper points to a practical middle path between brittle hand-coded automation and opaque learned policies: agent-generated computation graphs that can be simulated, inspected, revised, and then run on real variational automation tasks.
- Europe’s AI jobs map belongs at the workflow level OpenAI’s new EU workforce report is useful less as a prediction of job loss than as a planning tool for redesigning tasks, training, and procurement around occupations likely to face automation, growth, or workflow change across Europe’s labor market over the next few budget cycles.
- Ford's Layoff Rebound: What the Ford AI Story Should Teach Operators A widely shared claim that Ford replaced workers with AI and reversed course raises a real question for operators: where automation actually breaks, why rehiring happens, and how to deploy AI without paying for the lesson twice.
- Entity gap patching: the pSEO maintenance loop most teams skip A working note on using LLM entity extraction to keep programmatic SEO pages competitive without rewriting them. The job isn't generating more pages, it's patching semantic gaps in the pages you already have before rankings decay.
- Internal link audits on a 4,000-page site, done in an afternoon A working note on how I rebuilt our legacy internal link audit process using Claude 3.5 Sonnet and a small Python script. What the model is good at, where it falls apart, and the prompt structure that made the output trustworthy enough to ship.
- Linear AI rollouts are already too slow for marketing teams The standard SaaS pilot-then-expand playbook doesn't survive contact with AI tooling. Here's how marketing operators should structure parallel, use-case-driven adoption instead, and the operational tradeoffs nobody talks about when you abandon the comfortable one-team-at-a-time approach.
- Auditing 400 Old Blog Posts With a Local RAG Pipeline A working note on using local RAG to flag outdated SEO claims, stale stats, and content gaps across a legacy blog archive, with notes on what worked, what broke, and where the manual review still has to happen.
- Content decay analysis works better as a Claude prompt than a dashboard Identifying which blog posts are quietly losing traffic is one of the most tedious SEO jobs in marketing. Pairing a GA4 export with Claude 3.5 Sonnet turns a half-day audit into a 20-minute conversation, and the output is more useful than any decay dashboard I've built.
- Cursor is the spreadsheet moment for marketing ops Coding agents like Cursor have crossed a threshold where a marketer with no engineering background can build the scrapers, reporting tools, and automation glue they used to wait months for. Here is what that actually changes for a digital marketing operator in 2026.
- The marketer's stack for building internal tools without a dev team Cursor plus Claude has quietly become a viable way for marketing operators to build the scrapers, dashboards, and automations they used to buy or beg for. Here's how I think about the stack, what it actually replaces, and where it still falls apart.
- The Lovable test: can a marketer ship a working internal tool before lunch? Conversational app builders like Lovable are pitched as a way to skip engineering queues. I want to talk about where that actually holds up for marketing operators, and the specific micro-tools worth building first.
- The 4-second budget that decides if your AI agent ships Real-time marketing agents live or die by latency. Here's how I'm thinking about the sub-4-second budget when moving an LLM from demo to production, and which optimizations actually move the needle versus which ones just sound clever in a stand-up.
- Opus 4 is the tone-matching model. Stop using it like a generalist. Legal AI teams figured out something most marketers haven't: Claude Opus is uniquely good at matching the voice of a specific document or person. Here's how to port that workflow into brand copywriting without burning through your API budget.
- Agent Success Rate is the only number that matters when a new model drops When a frontier model ships, vibes-based testing wastes the window. Here's how marketing operators can build an automated eval harness that turns 'feels smarter' into a measurable percentage jump on the workflows they actually ship.
- Marketers are still vibe-checking prompts. Frontier devs run evals before lunch. Frontier developers test new models with automated eval suites and track agent success rates as a percentage. Most marketers eyeball outputs and call it good. Here's how to port the eval mindset into a content or SEO workflow without a research team.
- Stop Vibe-Checking New Models. Build a 50-Prompt Eval Set Instead. Frontier developers run automated evals the moment a new model lands. Most marketers still open the chat window and eyeball the output. Here's how to close that gap with a simple benchmark set of past briefs, ad copy, and SEO tasks you can rerun in an afternoon.
- Splitting the agent loop from tool execution cut TTFT by 90% Anthropic's Managed Agents architecture separates reasoning from tool execution, which solves two problems marketing operators hit hard: securing API credentials for tools like HubSpot and WordPress, and the latency tax of spinning up containers per session.
- Dreaming Agents Could Finally End the Brand Voice Correction Loop Anthropic's new Dreaming feature lets Claude review its own memory logs during downtime to decide what to keep. For marketing operators stuck retyping the same brand voice corrections, this is the first credible path to an AI editor that actually learns.
- How a 400-line system prompt becomes 15 lines with Skills A walkthrough of progressive disclosure as a fix for prompt bloat in marketing agents, why stuffing brand rules and SEO logic into one system prompt degrades reasoning, and how modular Skills restore eval scores while cutting token costs.
- The 200K Token CSV Problem Has a One-Line Fix Most marketing agents balloon past 200K tokens because operators dump entire CSVs into context. Giving the model a bash primitive to write and run Python against the file locally cuts cost, latency, and hallucination in one move.
- Moats Died When Model Releases Got Weekly Proprietary AI tech is no longer a defensible advantage for marketing teams. The real edge sits in execution speed, rapid model swaps, and feedback loops tight enough to ship before the next release cycle resets the board.
- Why I Stopped Trusting Demo Videos for Agent Tools Agent demos look magical until you try to ship one. Here's what I've learned about the gap between a polished demo and a workflow that actually runs reliably on Monday morning when nobody's watching.
- Why I Stopped Trusting My Own Prompts (And Started Logging Them) Most marketers treat prompts like throwaway text. The ones getting real output are treating them like code, with versions, evals, and a quiet discipline about what actually moved the needle.
- When Your AI Agent Needs a Browser, Not an API Most agent failures come from assuming everything has a clean API. The real unlock for marketers is teaching agents to operate the same messy web tools we do, with the same hands a human would use.
- Why I Stopped Trusting AI Demos and Started Timing My Own Workflows Most AI tool demos collapse the moment you put them against a stopwatch and a real client deliverable. Here is how I now evaluate whether a tool actually saves time in a marketing workflow, and why most of them quietly cost more than they save.