Helprism

Picture a small SaaS team that merges six pull requests this week, half of them written mostly by an AI coding agent. On Thursday a customer asks how to export invoices, and the help article still shows the button that moved on Tuesday. Worse, the AI answer bot on the help center confidently quotes the old article.

That’s the problem Helprism solves. It’s a help center that updates itself when you ship: the agent that wrote the change also drafts the help article update, a person reviews it, and readers can see when each article was last checked.

I started in February with a simpler idea: a help center that looks good out of the box and takes minutes to set up. It has five themes, dark mode, seven languages, custom domains, an embeddable widget and an AI answer bot that cites your articles.

But there are a lot of good-looking help centers out there. What none of them did was keep up with how fast small teams ship now. So in September I changed the pitch and the product: connect your repo or your coding agent, and every merged change turns into a reviewed article update. Nothing gets published without a person approving it.

This is the screen where most of the product lives. Updates come in grouped by pull request. Each one shows a rendered diff of the article with the words that changed highlighted, so you review the article the way your customer will read it and not as raw markdown. You can accept, edit or reject, and rejecting asks for a reason, so the next draft gets better. If the article changed since the draft was made, it tells you before you overwrite anything.

Customers brand their help center with their own colors, fonts and logo. When I redesigned the default theme, I had to be sure nobody’s saved branding would shift, so I checked that existing help centers rendered pixel-identical before and after. Boring to test, but it’s the kind of thing that loses you a customer if you get it wrong.

Helprism has 336 commits and around 880 tests, and AI agents co-wrote almost all of the code. For the first six months I committed straight to main with no CI, which was fine with just me. It stops being fine when there are dozens of agents working at the same time.

So I built a small system for it. I plan a round of work, and it gets split into lanes, each one an agent working in its own copy of the repo. Agents can write code, run tests and take screenshots, but a hook blocks them from merging, pushing, deploying, touching secrets or publishing packages. Those steps are mine. In eight days that was 57 lanes and 49 pull requests, and a file of rules that grows every time something goes wrong, so the same mistake doesn’t happen twice.

The design side works the same way. There’s a design doc the agents have to follow, with the tokens, spacing, motion and the things never to do. Agents explore directions as design boards, I pick one, and that board becomes the spec for the build.

The newest part of Helprism is an agent that reads a code change and figures out which articles it affects. Before trusting it, I built a test setup: sets of real changes with the right answers labelled by one model, judged by another, graded by a third, and some held back so I couldn’t tune for them. I also wrote down what score it needed before I saw any results.

It didn’t reach the bar. It found too few of the articles that needed updating. So it launched as a quiet beta, with no accuracy claims anywhere on the site, and the review step stays in place until the numbers say otherwise. I’d rather ship something honest than a nice number on the landing page.