Captain

Open to the public; beta testing with 12 journalists

I built Captain, a personalised AI editorial workspace for journalists. Growing up, both my parents were journalists. I watched as they lost their jobs as media outlets switched to using Echo Box, an enterprise AI publishing platform. Echo Box charges a ridiculous amount for something that is essentially an LLM wrapper, and they only sell to large businesses. I decided to build a competitor for individual, freelance journalists, of which there are many. In the transition to new careers while continuing their writing practices, or in transition to completely freelance income, individual journalists no longer have an edge over these AI-powered enterprises. They spend 16+ hours a day just to produce a tenth of the content their competitors do, and at a significant loss of quality. This is infeasible, and Captain aims to solve this issue.

I built Captain by following the daily workflow of one journalist, then expanded it so that different writers can configure it around their own voice, interests, and reporting process. Captain takes a story from discovery to a finished and journalist-reviewed draft. It curates relevant headlines, gathers supporting sources, and produces articles, research documents, LinkedIn posts, and X posts. Drafts are put next to source links and review flags so the journalist can inspect the evidence, make changes, and approve everything personally before it is sent out for publication.

“Personalisation” is one of the most important things I've built. The biggest issue with AI writing slop is that it is so general it becomes painfully obvious. Personalizing an LLM to write like you requires significant experimentation with long prompts and different models (funnily enough, the more “frontier” a model is, the worse it gets at personalizable writing). To ensure their writing is personalized to them, the journalist begins their journey with Captain by completing an interview, providing writing samples and reviewing five trials before activating their profile. Captain also has a learning process that compares its generated drafts against the journalist's edits and proposes reusable writing preferences. Each proposed preference has a few examples and a separate writing test for the journalist to review before approving it.

One of my biggest lessons (and something I'm ALWAYS tweaking) is that copying someone's writing voice requires understanding a lot more about them than their vocabulary and usual sentence lengths. It involves what angles they focus on within their subjects of interest, how they structure their arguments, which sources they trust and how audacious they are with their claims. These considerations influence both the onboarding process and the way Captain researches and drafts.

Another lesson is that not every single edit a journalist makes should be used to fine-tune generation. They might have caught a wrong fact or noticed that this particular story needs a different tone, but that says nothing about their overall stylistic preferences. I designed the learning process to distinguish between these cases and to ultimately leave the decision to preserve a style change in the hands of the journalist.

A major challenge, aside from the million bugs I had to fix, was keeping Captain simple to use. Journalists have access to all these tools, but using them and then stringing them all together is ridiculously overwhelming. Captain had to remain user-friendly at all costs. It has to manage a lot of different sources, prompts, drafts, edits, versions and approvals, but the journalist needs to be able to click just one button to get to the next obvious step. To keep this super simple I kept returning to the daily workflow of the average journalist: choose a story, review the research, write up a draft, edit it, and send it out. It sounds straightforward and it is, but building Captain taught me how much work keeping our processes that simple really is.

Captain is open to the public, but before campaigning any harder I am beta testing with 12 journalists. My focus is on making this as useful as possible to the working habits of a variety of journalists. I get regular feedback and am trying to optimize voice fitting and the learning process as much as possible based on this feedback, to ensure everyone is paying for the best possible product I can deliver to them.