Three years after joining Carta as a content marketer, Lucy Hoyle became the company's first content engineer. Her remit has already expanded beyond content. She now works across product marketing, events, demand generation, and other teams to improve existing workflows and design new AI-native systems.
Her story shows what marketing engineering looks like inside a large company. Lucy helped support a rapid Claude rollout across Carta's roughly 1,800 employees, then used Profound's Marketing Engineer Hackathon to prototype a shared context layer for the marketing organization. That prototype is now informing a broader marketing operating system.
The central lesson is that successful AI adoption requires more than faster execution. Teams need shared context, deliberate constraints, reusable infrastructure, and marketers who can connect technical systems to business outcomes.
Start with the outcome, then redesign the system
Lucy's first instinct with AI was familiar: take an existing process and make it faster. That approach can remove manual work, but it preserves the assumptions and limitations of the original workflow. Carta's CEO, Henry Ward, pushed the company to think more broadly about what AI makes possible.
For Lucy, leverage means stepping back from a linear workflow, defining the desired outcome, and reverse engineering a new system. Claude becomes a thought partner in that process, suggesting routes the marketer may not have considered. The useful question is no longer how to complete the same task faster. It is what the team could accomplish that was previously too difficult, fragmented, or large in scope.
That shift also changes the role of a marketing engineer. Lucy began as a generalist on Carta's small European team, writing long-form content, gated assets, social copy, and emails. That breadth helped her see how audience messaging, product positioning, regional knowledge, and channel execution fit together. As her role became more technical, she could apply that systems view across specialized teams instead of optimizing one content lane in isolation.
Train for different levels of AI fluency
Carta set out to get approximately 1,800 employees using Claude Chat and Claude Code within three to six weeks. Lucy joined an early-adopter group of technical leads and business champions, represented marketing with three colleagues, and helped turn the group's early lessons into a broader curriculum.
The rollout showed that company-wide AI training cannot assume one starting point. Some employees were already using Claude through a command-line interface and needed practical sessions that pushed the limits of what they could build. Others had never used Claude and needed a foundation in prompting and skills.
That lesson extended beyond internal enablement. Carta operates in a regulated industry and serves customers ranging from startup founders to investors. Different audiences bring different levels of familiarity and risk tolerance to AI, so the same principle applies when introducing AI-powered products or workflows to customers: understand their baseline, then tailor the education and experience accordingly.
Treat token limits as a design constraint
Carta did not turn token usage into a leaderboard. Lucy says the company used monthly caps and later added Slack alerts when usage spiked, but the goal was to help employees make conscious choices rather than make experimentation feel monitored.
That constraint encouraged a more important habit: deciding whether a task needs an LLM at all. If the work follows a repeatable process and does not require a model's reasoning layer, a Python script may be cheaper, more predictable, and easier to maintain.
This is a practical operating rule for AI-native marketing teams. The most sophisticated system is not necessarily the one that uses a model at every step. Good architecture separates deterministic work from work that requires judgment, then spends tokens where the model can materially improve the result.
Build one shared source of marketing context
As Claude adoption spread across Carta, employees began building skills and plugins quickly. The team initially shared them through Google Drive, which created predictable problems. People had to download and install their own copies, updates did not propagate automatically, and context was often hard-coded into individual SKILL.md files. Different people could end up running similar skills against different versions of company knowledge.
At Profound's hackathon, Lucy narrowed this organizational problem into a working prototype. She built what she calls a marketing brain: a shared context layer in a Git repository that skills and scripts can reference. For the demo, she connected Profound data to Claude, used an agent to gather competitive intelligence, and generated product messaging documents from a defined template. The resulting context files flowed back into the repository.
Git gives the system version control, a secure place to store internal context, and a way for multiple people to work from the same source of truth. Carta's engineers also built an internal Git server after the company ran out of GitHub licenses. The intended feedback loop goes in both directions: Claude can pull approved information into a task, then ask whether new campaign learnings should be committed back to the shared context layer.
Retrieve only the context each task needs
A shared repository solves fragmentation, but it creates a second problem. If every skill loads the entire marketing brain into Claude, the model's context window fills quickly. Lucy learned through Carta's AI learning labs that supplying more information can degrade performance when relevant instructions and facts become harder to retain.
Her solution was a retrieval layer built with BM25, an open-source ranking method also used in Elasticsearch. When someone starts a conversation or runs a skill, the system identifies keywords in the request, matches them against the marketing brain, and returns only the relevant chunks. Claude receives focused context for the task instead of the entire repository.
This pattern turns context engineering into infrastructure. Individual marketers do not need to know which files to attach or how much background to include every time. The retrieval layer applies progressive disclosure automatically, while the shared repository remains broad enough to support many teams and use cases.
The hackathon prototype has since become part of a larger plan. Lucy and several colleagues spent two to three days consolidating different ideas into what they call Carta's marketing OS. The team is now mapping what information each marketing function consumes and produces, curating shared skills through a committee, and deciding whether to start with one campaign use case or ingest context more broadly.
Lead with impact when presenting technical work
Building the system is only part of Lucy's role. Getting support requires translating it for managers and executives who care first about the business outcome. Her advice is to begin with the impact, the reason the project matters, the cost of inaction, and the expected organizational value. Technical details should follow only when the audience wants them.
That principle helped Lucy and her colleagues present the marketing OS at a company all-hands and communicate regularly with Carta's CMO. It also provides a path for marketers who want to move into engineering work. Build something useful, show the result, and explain it in the language of the audience whose support you need.
Lucy's own transition is evidence for that path. She did not begin as an engineer or try to master every technical layer at once. She learned enough Git, scripting, Claude, and systems design to solve real marketing problems, while keeping the audience knowledge and business judgment she developed as a content marketer.
Listen to the full conversation
Listen to the full episode for Lucy's account of moving from content marketer to Carta's first content engineer, supporting Claude adoption across 1,800 employees, prototyping the marketing brain at Profound's hackathon, using BM25 to prevent context rot, and turning that work into a broader marketing operating system.
