How We Use AI to Produce the Daily Briefing

The Daily AI Briefing (our newsletter) is researched and drafted by an automated Claude workflow, then reviewed and edited by a human editor before it publishes. This page describes exactly how that works: what the workflow does, what it does not do, the checks every issue passes before a human ever sees it, and how to report an error.

We're an AI advisory firm. We tell clients to disclose AI use, keep a human in the loop at defined checkpoints, and build verification gates before anything ships. This newsletter is a working demonstration of all three, running in public, multiple times a day.

What the workflow does

A few times each day, a Claude workflow researches current AI news and drafts a newsletter issue: a short lead analysis and a handful of linked, attributed news items. It writes the draft as a file in our content repository, runs it through a set of automated checks, and if those checks pass, opens a pull request for a human editor to review.

The workflow never publishes directly. There is no path from the workflow to the live site that skips a pull request and a human review.

What the workflow does not do

The workflow does not fabricate statistics, quotes, or company names. Every item in an issue links to a real, reachable source. It does not claim first-person experience: you will not see "we spoke with" or "in our work with clients" in the automated draft, because the workflow has no clients to speak with. First-person client experience appears only in insights and research papers written by a person.

The workflow does not publish on a schedule regardless of output. If a run finds nothing new and relevant worth reporting, it exits without opening a pull request. A missed run is a normal, expected outcome, not a failure.

The checks every issue passes

Before a draft ever reaches a human reviewer, it runs through automated checks that block on:

• A source URL for every item that actually resolves, checked live at validation time
• No item repeating a story already covered in the last 30 issues
• A named publication for every source, and a substantive lead analysis
• The same style rules that apply to every page on this site: no em dashes, no banned words or filler phrases, no fabricated claims
• Frontmatter and file consistency: the issue's date, number, and metadata all agree with each other

Any failure blocks the pull request from being opened. These checks catch dead links, duplicate stories, and style violations. They cannot catch everything, which is exactly what human review is for.

What a human reviews

Every issue is reviewed and edited by a person before it publishes. No exceptions: our systems enforce this by blocking the merge of any issue that hasn't been marked reviewed. The reviewer checks:

• Whether the linked source actually supports the claim attached to it
• Whether the lead analysis is a real judgment, not a restated headline
• Whether it's the right story for what matters to our audience that week
• Whether it sounds like us
• Whether anything in the draft would be embarrassing to publish under our name

The reviewer edits freely. A pull request from the workflow is a draft, not a proposal to accept as written.

How every issue discloses this

Every newsletter issue carries a short line near the byline stating that it was researched and drafted by an automated workflow and reviewed by a human editor before publication. The same disclosure is embedded in the page's structured data and in our RSS feed, so it travels with the content wherever it's read.

How to report an error

If you find a factual error, a broken link, or anything else wrong in an issue, email hello@roai.us with the issue date and what's wrong. We'll investigate and correct it.