A group of tech freelancers had a side project they were serious about. They wrote under a shared brand with individual bylines. Their driving principle was refreshing: don't repeat what's already in the documentation. Go deeper, find the angles that matter, write for the reader who already knows the basics. The production process hadn't kept up with the ambition.

We started by asking what each phase of the process was actually trying to accomplish. What we found was that every phase had a fundamentally different character — some needed consistency and reliability, others needed flexibility and room to explore. Building one system that treated all eight phases the same would have served none of them well. So we didn't.

Ideas arrive at inconvenient moments. The repo needed to capture them as frictionlessly as possible — a title, an optional description, an optional URL, installable on a phone. The detail that made it work: users could share content directly from anywhere. Reading something in the New York Times? Share it to the repo, add context, done. Structure just enough to make ideas retrievable. Intelligence in removing every barrier between inspiration and capture.

A single action stubbed the article's entire folder structure on shared Dropbox and announced the commitment to the group. A new card appeared on the shared Trello board. Structure handled the setup. The social signal of a new card handled the accountability.

The first version was simple: a structured prompt sent to an AI API, results captured and filed. It worked. It also felt insufficient almost immediately. We rebuilt research as a multi-phase process — early prompts for broad exploration, later prompts going deeper into areas the author had chosen to pursue, side research for tangents. We used both the Claude API for online work and a local Ollama instance for research the group wanted to keep internal. A custom Trello power-up previewed all research files directly on the card — summary, raw markdown, and rendered preview in one tabbed interface.

Tech articles at this level require hands-on work with the technology. We made saving sandbox projects to GitHub straightforward, with repo links automatically added to article metadata. An AI assist tool similar to the research tool helped authors think through what they were building and why.

Tools designed around specific needs — turning bullet points into structure, finding the right entry point for a complex idea. None of them were "write the article." That distinction mattered to the group and it mattered to us. The tools assisted the author's thinking. They didn't substitute for it.

Two tracks: human feedback from other collective members, and AI-assisted review using specifically constructed prompts — grammar, technical accuracy, accessibility to a novice reader. Each produced a structured response, saved and visible on the Trello card.

A tool moved the article from the file structure into Sanity Studio. Once approved, the publication date was written to the article's metadata and added to the Trello card as a due date. Trello's own automations handled the rest. The article published itself.

Eight phases. Each one designed around what that stage actually needed. The Tech Girls ended up with something that felt less like a system imposed on their process and more like their process, finally given the support it deserved.