$ cat METHODOLOGY.md # no secrets in how this is made
RepoJournal is an AI-assisted publication, and we would rather explain exactly how that works than let you guess. This page is the whole method — what the pipeline does, what a human does, and what happens when we get something wrong.
Every briefing is built exclusively from public GitHub activity: commits, pull requests, releases, and security advisories. Nothing is scraped from other publications and nothing is invented. Every item's heading is the stored title of the underlying commit, PR, or release — the model never writes headlines — and links straight to that source, so every claim can be checked in one click. Statistics (commit counts, contributor tallies, release figures) are computed from the raw activity, never estimated by a model.
A language model writes the connective prose: the one-to-two-sentence blurbs beneath each item heading, the day's lede, and the weekly review's through-lines. It works under hard rules we test in code: it may only reference the stories it was shown, it may not invent version numbers or figures, quotes must be attributed, and "action items" survive only when a story creates a concrete obligation — a security patch, a breaking change — never as filler. When a day is genuinely quiet, the briefing says so instead of inflating routine work.
AI does not choose what exists: story selection is a scoring pipeline over the raw activity, headlines are the sources' own words, and contributor credits come from commit metadata.
The weekly and monthly reviews — the pages we consider our editorial flagship — get a standing human editorial pass every week: reading for errors, cutting anything that doesn't hold up, and feeding corrections back into the pipeline's rules. Our editors publish pseudonymously; the accountability is organisational and it is real — every correction lands publicly on the page it amends.
If a briefing misstates what shipped, we want to know: hello@repojournal.com, or reply to any email we send — replies reach a human. Corrections are made on the page, not quietly deleted.
The daily wires exist for subscribers and are deliberately kept out of search indexes; the pages we stand behind for reference are the per-project front pages, the weekly and monthly reviews, the topic hubs, and the Breaking Changes Index. In that index, security advisories publish automatically from GitHub's own advisory data; editorial entries — deprecations, removals, breaking releases — are human-reviewed before publication, and everything is covered by the weekly editorial pass.
More context: how RepoJournal works · privacy