This case study is about how content gets produced at scale—the AI skills and drafting tools a whole product org writes with. Where the standards live and how they get enforced (the design system, the PR checks) is its own case study: Copy, shipped. Same standards, different problem.

One set of standards, six packaged judgments, delivered where people already work.

context

In 2025, our writing team shrank to one person: me. The requests didn't shrink with it. I was covering UI copy and documentation across nearly every product area, and the honest options were: become the bottleneck, let quality slide, or change what content quality depends on.

I chose the third. Over the following year, I built a content AI system (not a chatbot—an architecture) that encodes our standards once and delivers them everywhere people actually work.

the system

Three layers, deliberately separated so each can update on its own:

Standards: the source of truth. Voice, tone, terminology, component-level copy guidance, and documentation conventions, maintained as one canonical context pack instead of scattered wiki pages.
Skills: the judgment layer. Six custom AI skills built on those standards—three for documentation (article generation, article review, docs triage) and three for UX writing (copy design, naming and terminology, prompt design for AI features). Each one packages more than rules. It packages how I'd apply them: what questions to ask, which tradeoffs to weigh, what good looks like with examples.

Screenshot of the Claude skills directory showing the organization's shared skills, with six content design skills highlighted: copy-design, prompt-design, article-generation, docs-triage, naming-terminology, and article-review, each with a description and download count.

Live in the org's shared skills directory. (The Bufo emoji creator isn’t mine, but I respect it.)

Delivery: meeting people where they work. The same standards power different surfaces for different audiences. Claude Projects for anyone who needs polished product content—PMs, designers, and engineers, but also marketing or customer-facing teams. A context pack in the engineering ai-tools repo, so anyone prototyping in Cursor—designers and PMs especially—can invoke /ux-writing or /docs-generation from any repo and start closer to right. The automated copy reviewer on every frontend PR, so engineers get the same standards without leaving their workflow—ideally by running /code-review themselves before opening the PR, with the post-merge check as the last line rather than the first. And the release-docs pipeline, where PMs approve what the standards already drafted.

Plus governance: a maintenance page documenting how the pieces sync, download links, update procedures, and a review cadence. (An AI system without upkeep is just future drift.)

adoption, with receipts

Alt: Slack announcement from Maddy launching two custom GPTs for the team: a Product Content Assistant for naming, UI copy, and guides, and a Docs Assistant for drafting Knowledge Base content, with a warning that the tools support early drafts only and the writing team should review before dev handoff. The post has over two dozen positive reactions.

The first launch: two custom GPTs and a guardrail—early drafts only, writing team in before dev handoff.

Slack announcement from Maddy launching the next generation of the content AI system: six Claude skills (three docs, three UX writing), a content context pack in the ai-tools repo for Cursor users, and updated documentation, with install instructions and a request for feedback.

Eight months later, the second launch: six Claude skills and a context pack in the engineering repo. Same standards, maturing delivery.

By Pendo's own analytics: the ChatGPT version of the Product Content Assistant reached 117 colleagues—about 10% of the entire company—across ~4,000 sessions averaging 1h 27m each, with usage spanning every department. Its Claude successors are newer and already tracking dozens more. And all of these numbers undercount—Pendo only sees browser usage, not the desktop apps, where most users shifted.

Pendo analytics dashboard for the ChatGPT Product Content Assistant showing 117 visitors, 100% account adoption, about 4,000 page views averaging 1 hour and 26 minutes per visitor, and a weekly usage chart peaking in late 2025 then declining as usage shifted to other tools.

14 months of usage—the dip isn't abandonment; the org moved to Claude and desktop apps this dashboard can't see.

Our co-founder and CPO asked for a three-week pilot on one product area. I built the pipeline for all product areas instead: it scans every merged change, drafts and grades the help-article update, routes it to an approver, publishes, and logs every run. Content designers tested it for three weeks; on October 15, 2026, I handed it to PMs—five to start, every PM by year-end—so I could move to Novus full-time.
In its first two weeks, the pipeline published 33 updates across nine product areas, most with the original drafts, every one scored against our rubric before a human saw it—and when a write truncated a 68k-character article, the retro produced three new safety rules the same day. Release docs went from a writing queue to a review step.
A senior staff product designer called the context packs and the GPTs built on them "some of the most practical and forward-thinking tools I've seen"—noting they'd "quickly become part of how PMs and designers draft and refine language."