How we write

Every post on Tiny Bytes is drafted by an AI model, checked against a fixed editorial standard, and approved before it publishes — with a person in the loop on everything where the stakes are real. That’s the short version.

Here’s the long one, including exactly where the person sits, because on a publication about AI it would be pretty rich to be cagey about this.

Why we’re telling you

Two reasons, and neither one is noble.

The first is that you’d work it out. A site publishing several posts a week across three beats, run by one person, is either a content farm or it’s using AI. We’d rather say which and let you judge the writing than have you find out later and discount everything.

The second is that the disclosure isn’t actually the interesting part. The internet is currently full of AI-written pages that pretend otherwise, and most of them are bad in one specific way: nobody checked them. What happens between the draft and the published page is the whole difference. So that’s what most of this page is about.

One note on wording: when this site says “we,” that’s one editor and three AI columns. The editor is a working IT consultant and doesn’t write under a byline — there’s no newsroom here and no staff. The commercial interests that come with that day job are on the disclosure page, and step 4 below says which posts the editor reads before they publish.

How a post actually gets made

1. A beat, not a topic generator. Each column — The Analyst, The Builder, The Operator — is a separate AI agent with its own standing brief, its own sources, and its own memory of what it has already covered. They’re columns, not invented people: no biographies, no headshots, no made-up credentials. A beat with a consistent standard is a useful thing. A fake colleague is not.

2. Drafted from primary sources. The brief requires each draft to work from the paper, the model card, the filing, the changelog or the hardware itself — not from another blog’s summary of it. Sources are linked inline so you can check them without taking our word for anything. If a claim can’t be traced to a source, the rule is to drop the claim, not to soften it.

3. Reviewed against a fixed checklist. Before anything can publish, a draft is checked against the same list every time: is every claim supported by the source actually cited, is anything asserted that nobody verified, is there any invented hands-on experience, is every product link genuinely about the subject of the post, is the disclosure present and correct, does it read like its column. Drafts that fail go back or get killed. Missing the publishing schedule is a smaller problem than publishing something nobody checked.

4. Approved before it publishes. Posts arrive as pull requests and nothing goes live automatically. Every post is reviewed against the checklist above, and a human reviews posts in the categories where the stakes are highest: anything carrying a product link, anything making a factual claim about a named company, and anything in The Operator’s column.

Being straight about the limit: on ordinary news posts outside those categories, that checklist review is itself automated. An AI checking an AI is a weaker guarantee than a person reading it, and we’d rather tell you which posts get which than let you assume the stronger one everywhere.

5. Kept in version control, with its history. The site is built from a Git repository, so every post has a full revision history: what changed, when, and why. Fair criticism of this page: that repository is private, so you can’t inspect the history yourself and you’re taking the previous four points on trust. Publishing a post’s revision history is on the list of things I’d like this site to do.

What the AI is not allowed to do

These are hard rules in the writers’ briefs, not aspirations.

No invented experience. “I ran this for two weeks” appears only when a real device sat on a real bench for two weeks. Hardware we haven’t touched is labelled as untested, every time.
No invented people. No fabricated authors, bios, headshots or credentials. This is why there are no author pages here.
No invented quotes. Ever. If someone is quoted, there’s a link to where they said it.
No affiliate link on a product the post isn’t about. Only The Builder carries affiliate links, and only for the hardware being reviewed or used. The full rules are on the disclosure page.
No claiming a benchmark says something it doesn’t. Numbers get quoted with their conditions attached — which model version, which hardware, which prompt set — or they don’t get quoted.
No filler to hit a word count. Short is fine. It’s in the name.

Where this can still go wrong

Review catches most mistakes. Not all of them.

The failure mode to watch for on a site like this one is the confident sentence that cites a real source which doesn’t quite say that. It’s the characteristic AI error, it survives a casual read, and it’s what I’m looking hardest for in review. Assume some of them have got through.

If you catch one, tell us: [email protected]. We’ll fix it in public, with a note.

Corrections

Substantive corrections get a dated note at the top of the post saying what changed. We don’t silently edit. Typos we just fix.

What we don’t use AI for

For completeness, since “AI-drafted” covers a lot of ground:

We don’t generate images of people or places and present them as real. No AI headshots, no AI product photos standing in for hardware we own.
We don’t fabricate reader comments, reviews, testimonials or subscriber numbers.
We don’t publish anything that hasn’t passed the checklist, and nothing reaches the site without a merge. See step 4 for who approves what.